Claude Blog 采集 (2026-05-18)¶

共采集 10 篇文章

📋 文章索引¶

Building AI agents for healthcare and life sciences - Oct 30, 2025 (评分: 9.5)
How three YC startups built their companies with Claude Code - Nov 17, 2025 (评分: 9.5)
Building multi-agent systems: When and how to use them - Jan 23, 2026 (评分: 9.5)
How Carta Healthcare gets AI to reason like a clinical abstractor - Apr 08, 2026 (评分: 9.5)
Claude Code and new admin controls for business plans - Aug 20, 2025 (评分: 9.5)
Long context prompting for Claude 2.1 - Dec 06, 2023 (评分: 9.0)
Claude 2 on Amazon Bedrock - Aug 23, 2023 (评分: 9.0)
Claude and Slack - Oct 01, 2025 (评分: 9.0)
Claude API skill now in CodeRabbit, JetBrains, Resolve AI, and Warp - Apr 29, 2026 (评分: 9.0)
Claude now creates interactive charts, diagrams and visualizations - Mar 12, 2026 (评分: 8.2)

Building AI agents for healthcare and life sciences¶

来源: Claude Blog 发布日期: Oct 30, 2025 采集时间: 2026-05-18 价值评分: 9.5/10 正文字数: ~15397 字符

摘要¶

AI agents are delivering measurable results in healthcare and life sciences while navigating the industry's unique regulatory complexity and safety standards. Here's how.

正文内容¶

AI agents are delivering measurable results in healthcare and life sciences while navigating the industry's unique regulatory complexity and safety standards. Here's how.

In healthcare and life sciences, AI is proving its worth where it matters most: real-world patient outcomes. AI agents are delivering measurable efficiency gains: 16,000 hours of annual research time saved at Pfizer and clinical documentation study delivered in hours, not weeks at Novo Nordisk . For technical leaders evaluating their agentic investments, these production implementations demonstrate concrete ROI. The question isn't whether to build agents—it's how to architect systems that meet healthcare's unique regulatory requirements and clinical safety standards while delivering real results. From analysis to action: the agent advantage Agents represent a fundamental shift from AI that requires constant human input to systems that reason through problems, plan multi-step approaches, and execute tasks. This shift is especially impactful in the healthcare world, where data often lives in fragmented systems that don't talk to each other, making it harder to see the complete picture of a patient's health. Access to these systems let alone the expertise to understand one if not all of the data types is prohibitive at worst and time consuming at best. Agentic systems can understand clinical context, ingest information from multiple unrelated sources, process multiple kinds of data (images, text, audio, etc.), and apply all of these capabilities to taking meaningful actions within healthcare workflows. What does this look like in practice? Instead of a clinician manually pulling data from five different systems, reviewing it, and then updating a care plan, an agent can monitor patient vitals across those systems, recognize concerning patterns, draft updated care recommendations based on current guidelines, and route them to the right clinician for approval. The agent handles the coordination and analysis; the clinician makes the final decision. This shift from traditional AI to AI agents is particularly significant for healthcare because it tackles the process completion problem. Healthcare workflows don't just need information, they need actions taken across multiple systems to actually complete care delivery. Agents can bridge those gaps. Production implementations delivering measurable results AI-powered healthcare agents are already delivering real-world results in areas such as hands-on medical care, healthcare organization administrative loads, and researching new drug treatments. Research acceleration and drug development Pfizer reduced annual research time by 16,000 hours using Claude to handle literature review, data synthesis, and documentation, freeing researchers to focus on scientific problem-solving rather than tedious workflow management. For organizations facing pressure to accelerate drug development timelines while managing costs, this productivity gain demonstrates measurable ROI. Administrative workflow automation Novo Nordisk automated clinical study report generation using Claude, reducing documentation production from 10+ weeks to 10 minutes. The pharmaceutical company built NovoScribe, an AI-powered platform using Claude Code and MongoDB Atlas, to handle regulatory documents that previously required entire departments. Clinical study reports—comprehensive trial summaries running up to 300 pages—now generate automatically while maintaining regulatory compliance standards. Staff writers who previously averaged 2.3 reports annually can now produce documentation in minutes rather than months. Healthcare-specific considerations Despite the obvious benefits, healthcare agent implementations face unique technical challenges: data fragmentation across incompatible systems, strict regulatory requirements, and direct impact on patient safety. These constraints require different architecture decisions than typical enterprise deployments. The combination of complex medical contexts, regulatory requirements, and direct impact on patient outcomes creates an implementation environment where thoroughness trumps speed. In this section we will address three distinct challenges and how to address them to get the most from your agentic solutions. Data fragmentation and interoperability Data fragmentation across healthcare ecosystems creates significant integration challenges. Clinical departments operate specialized systems—radiology maintains image repositories, labs store results in separate databases, pharmacy records live elsewhere entirely. Legacy medical devices compound these problems, many developed decades before modern API integration patterns existed. Integration complexity often shows up in: EHR vendor incompatibilities (Epic, Cerner, AllScripts) Departmental data silos requiring cross-system orchestration Legacy medical device integration challenges Real-time synchronization needs for time-sensitive clinical decisions As a result, successful integration requires three key decisions: 1. Connectivity approach: Direct integration with existing systems Custom connectors (APIs or MCP) Middleware to bridge communication gaps 2. Data formatting: Standardized ingestion processes for varied data types Format conversion between incompatible systems Handling unstructured clinical text vs. structured data 3. Synchronization requirements: Real-time access for time-sensitive decisions Latency tolerance based on clinical urgency Batch processing for non-urgent workflows" Regulatory and compliance requirements Regulatory compliance has to be part of every organization’s agent architecture from day one, not as an afterthought. Almost every country has baseline requirements around data privacy and comprehensive audit trails that fundamentally shape your system's architecture. If you don't build with those requirements in mind, you will be forced to rebuild sooner rather than later, possibly after facing steep fines. Some regulatory considerations that need to be part of your agent architecture include: HIPAA compliance for AI data processing workflows Evidence-based validation of patient outcome improvements Documentation requirements for AI decision audit trails Generic enterprise AI solutions don't meet healthcare regulatory requirements. HIPAA's recent cybersecurity guidelines require comprehensive observability and accountability for AI systems processing PHI, while the EU AI Act classifies most healthcare AI solutions as high-risk systems. Your healthcare organization needs robust risk management systems, clear data governance ensuring bias-free training data, detailed technical documentation, and human oversight mechanisms that go beyond checkbox compliance. Maintaining human authority and clinical oversight Agent autonomy delivers efficiency gains, but clinical decisions require human authority. Implementation architecture must enforce this boundary through: Transparent reasoning that clinicians can understand and validate Clear escalation pathways for complex or ambiguous cases Override capabilities allowing clinicians to reject AI recommendations Fail-safe defaults that prioritize patient safety over operational efficiency The goal is creating AI agents that make excellent clinicians even better, not replacing the human expertise that patients depend on for compassionate, nuanced care. When implemented thoughtfully, these systems enhance diagnostic accuracy by spotting patterns across massive datasets, apply evidence-based guidelines consistently during hectic clinical situations, and alert doctors to subtle signs of patient deterioration before they become emergencies. Throughout these applications, clinical authority remains intact while administrative burden decreases, allowing more focus on direct patient care. Implementation strategy Production implementations at Pfizer and Novo Nordisk demonstrate specific efficiency gains: 16,000 hours of annual research time saved, automated processing of thousands of daily referrals, and 40-50% improvements in quality control workflows. How do you build and deploy agents that deliver similar results while meeting regulatory requirements and clinical safety standards? Where to start The best agent initiatives begin by targeting the things that everyone already agrees need fixing. Clear metrics make all the difference here, because they show whether the solution actually works and help build momentum for wider adoption. Pick your battles Initial implementations should target high-visibility or widely felt problems with clear metrics. Documentation efficiency stands out as the most promising starting point—organizations see measurable results quickly, clinical staff feel the immediate impact, and voice-based documentation delivers especially rapid wins that build organizational confidence for broader deployments. The opportunity extends beyond simple transcription. AI agents can condense lengthy patient encounters into concise summaries, extract relevant data points from unstructured clinical notes, and convert information between incompatible system formats. Patient engagement and communication are another promising area for initial implementation, especially for routine administrative tasks and basic patient inquiries. Letting an AI agent take over tedious or repetitive work not only provides immediate operational benefits; it also helps build overall trust and confidence that agents can be a boon to other areas of your organization. Be thoughtful and prioritize compliance Diagnostic support delivers significant value but requires thoughtful planning due to regulatory requirements and workflow integration complexity. Start with lower-risk applications that keep humans in the loop: Abnormal lab value flagging: Systems highlight concerning results for clinician review Drug interaction checking: Automated recommendations that pharmacists validate Clinical guideline reminders: Evidence-based suggestions during care planning These applications provide immediate value, integrate into existing workflows without major disruption, and build team experience with agent systems before tackling higher-stakes use cases. Building momentum Success with initial implementations opens the door to enterprise-wide capabilities. The key is moving from point solutions to shared infrastructure that serves multiple departments. Core capabilities Build shared infrastructure that serves multiple departments rather than isolated point solutions. This approach reduces redundant development costs, accelerates time-to-value for new use cases, and ensures systems evolve alongside advancing technology capabilities. For example, a single natural language processing engine that understands clinical terminology can power documentation automation, patient communication systems, and clinical note analysis across your organization. Rather than three separate NLP investments—one for documentation, one for patient communications, one for clinical decision support—you can build institutional knowledge in one area and deploy it across multiple use cases. The same principle applies to unified data integration platforms: build once, deploy across departments, and maintain centralized expertise as technology advances. Building trust through transparency Agents interact with both your workforce and your patients, and each group will respond differently to AI-assisted processes. Trust-building matters as much as technical deployment. For patients, transparency matters. Make it clear when they're interacting with an AI agent versus a human, explain what the agent can and cannot do, and provide straightforward pathways to human specialists when needed. This clarity builds confidence and encourages broader adoption of AI-assisted services. Internal adoption follows similar principles. Your organization already has change management processes for new clinical systems. Apply them here. Staff need to understand how agents work, when to trust their recommendations, and how to escalate concerns. Present agents as partners that handle repetitive coordination work, allowing clinical teams to focus on what they do best: complex medical decisions and patient interaction. Tackling the bigger challenges At this stage, your organization has built core capabilities, demonstrated measurable wins, and developed teams with hands-on agent experience. The lessons learned from early implementations provide the foundation for more complex deployments. The observability and human-in-the-loop mechanisms you built for simpler use cases become even more critical as complexity increases. More than ever, your implementation needs: Comprehensive audit trails that track every agent decision and data source used, enabling clinical review and regulatory compliance Real-time monitoring systems that detect when agents encounter edge cases or uncertain scenarios requiring human judgment Escalation protocols that automatically route complex cases to appropriate clinical experts based on clear criteria Performance metrics that measure not just technical accuracy but clinical outcomes and workflow integration success Driving AI transformation at scale For technical leaders evaluating agent investments, production implementations at Pfizer, Novo Nordisk, and other organizations demonstrate measurable efficiency gains across research, administrative workflows, quality control, and regulatory documentation.: Success requires thoughtful implementation that balances technological capability with the unique challenges to the healthcare industry. This approach delivers quick wins that build confidence while establishing the technical foundation needed for evolving into more sophisticated agent initiatives. Evaluate whether your workflows have sufficient complexity to warrant agent architectur e. As a general rule of thumb, some organizations achieve better ROI with simpler automation solutions that address their specific problems more efficiently than full agent implementations. The path forward demands partnership between technology and clinical teams to ensure AI agents enhance, rather than replace, human judgment. Healthcare leaders who embrace this collaborative approach, prioritize patient safety through robust testing and escalation pathways, and build modular systems that evolve to keep pace advancing AI capabilities will lead the way in creating strong, sustainable, and incredibly valuable systems for their organizations. Production implementations prove the gains are achievable: measurable efficiency improvements across research, administrative workflows, and quality control. Your path forward is architecting systems that capture these benefits while maintaining the rigorous regulatory and clinical safety standards healthcare demands.

Accelerate your enterprise AI transformation with proven strategies from Anthropic's customers in HCLS.

Learn how life sciences organizations are

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

How three YC startups built their companies with Claude Code¶

来源: Claude Blog 发布日期: Nov 17, 2025 采集时间: 2026-05-18 价值评分: 9.5/10 正文字数: ~14975 字符

摘要¶

Learn how three YC startups use Claude Code to ship products faster, win government contracts, and scale AI-powered platforms with agentic coding workflows.

正文内容¶

From non-technical founders winning government contracts to solo devs building at scale, here’s how agentic coding is re-writing the startup playbook.

Y Combinator , a startup accelerator, has launched over 5,000 companies that have a combined valuation of over $800B since 2005, including household names like Airbnb, Stripe, and DoorDash. Today, agentic coding tools like Claude Code are fundamentally changing how YC startups build and scale. Founders can now ship products directly from the terminal, compressing development cycles from weeks to hours and enabling even non-technical founders to compete with established players from day one. We spoke with three YC startups who demonstrate this transformation in action: HumanLayer (F24) built their entire platform and pioneered context engineering practices with Claude Code Ambral (W25) is scaling AI-powered account management with sophisticated sub-agent powered workflows built with Claude Code and Claude Agents SDK Vulcan Technologies (S25) is using Claude Code to tackle regulatory complexity for the government and industry Let’s dive in. HumanLayer: From SQL agents to scaling AI-first engineering teams Dexter Horthy was building autonomous AI agents to manage SQL warehouses when he noticed a fundamental (but understandable) challenge to agentic adoption: companies weren’t comfortable giving AI applications unsupervised access to sensitive operations like dropping database tables. The product pivot that started it all That realization became HumanLayer's core insight: often the most useful functions in any software are also the most risky, especially for non-deterministic LLM-powered systems. "Our MVP was an agent that would coordinate with humans in Slack and could do basic cleanup, like dropping any table that hadn't been queried in 90+ days," Horthy explained. "We weren't comfortable with an AI application running raw SQL unsupervised, so we wired in some basic human approval steps." In August 2024, Horthy built an MVP, demoed it to different startups across SF, and had his first paying customers. This progress landed HumanLayer in the YC F24 batch, and the team went all in on providing an API and SDK that lets AI agents contact humans for feedback, input, and approvals across Slack, email, SMS, and other channels. Through Q1 2025, the HumanLayer team conducted extensive customer discovery, talking to dozens of engineering teams building AI agents and realized there was a gap in the agentic development loop they hadn’t accounted for. "Every team had rolled their own agent architecture," Horthy explained. "We realized we couldn't just build a better API–we needed to help establish the patterns and principles that would let the ecosystem mature." This led Horthy to document their learnings in " 12-Factor Agents: Patterns of reliable LLM applications ." Published in April 2025, the guide went viral and synthesized their experience building production agent systems and highlighted best practices for the emergent discipline of context engineering. Building everything with Claude Code With these learnings in hand, the HumanLayer team started exploring alternative product ideas and pivot angles. When Anthropic launched Claude Code, Horthy and his team were already strong proponents of Claude models for coding. They immediately began using it to build these experiments. "We just wrote everything with Claude Code," Horthy said. "When the Claude Agent SDK launched with Opus 4 and Sonnet 4, enabling headless agent execution, we knew this was going to be a big deal." After months of refining their Claude Code workflows internally, Horthy began sharing them with close founder friends. "The moment that told me we needed to go all in on this was an all-day pairing session with Vaibhav from BoundaryML (YC W23)," Horthy recalled. "Vaibhav was skeptical at first, but after we spent 7 hours shipping what would normally be 1-2 weeks of work, he was sold. I realized this workflow could work for other teams and other codebases." Building CodeLayer: Scaling AI-first engineering Today, HumanLayer's product CodeLayer helps teams run multiple Claude agent sessions in parallel using worktrees and remote cloud workers. They've discovered a critical pattern: once an engineer masters Claude Code, their productivity gains are so substantial that the real challenge becomes organizational—scaling these workflows across entire teams. "Once you have multiple people on your team shipping AI-written code, you have a completely different type of problem," Horthy explained. "It's a communication, collaboration, tooling, and management problem. You have to rewire everything about how your team builds software." Since the start of Q4 2025, HumanLayer has closed several large pilots across engineering teams of all sizes to deploy these tools and workflows, all built with Claude Code. Ambral: Building production systems with subagents Jack Stettner and Sam Brickman founded Ambral to solve a problem familiar to every B2B startup founder and CRO: as companies scale, the founder-level customer intimacy that drives early growth becomes impossible to maintain. Transforming account management with the Claude Agent SDK Whether at early companies experiencing hyper growth or at established enterprise companies, account managers routinely juggle 50 to 100 accounts simultaneously. "You can't give an effective account management experience with 1/50th of someone's attention," Stettner explained. Customer context that once fit in a founder's head scatters across systems, logs, Slack messages, meeting transcripts, and product usage data. Ambral synthesizes signals from customer activity and interactions into AI-powered models of every account. The system pinpoints who needs attention and why, autonomously driving or recommending expansions while catching early signs of dissatisfaction to prevent churn. "We're trying to provide the experience of every customer having a one-to-one account manager," Stettner said. As CTO and sole engineer at this young startup, Stettner relies heavily on Claude Code for development and Claude's Agent SDK to power the product itself. The technical architecture reflects sophisticated understanding of how to extract maximum value from different Claude models. Delegated workflow: Opus for thinking, Sonnet for building, and subagents all around Stettner has adopted a precise workflow that leverages the strengths of different Claude models in conjunction with subagents: "I use Opus 4.1 to do research and planning. Sonnet 4.5 has been absolutely killer in terms of being able to then go and implement these plans that I create in Markdown," Stettner explained. His development process follows three discrete phases: Research phase (Opus 4.1) : Perform deep research on whatever background is needed for a feature implementation. "I think the most important thing is doing research before you plan," Stettner emphasized. "Have Claude do research for you and create a large, long research document." He uses a series of subagents to research multiple areas of the codebase in parallel. Planning phase (Opus 4.1) : Create a plan with discrete phases on how to implement the feature. "I'll have Opus create a plan with phases, discrete phases on how to actually go about implementing it, and I'll go revise that plan. Maybe I'll chat with Opus about questions about certain details, or I'll manually update this markdown file." Implementation phase (Sonnet 4.5) : Execute each phase of the plan systematically. "Then I'll use Sonnet 4.5 to go and implement each phase." This approach prevailed over the other workflows Stettner tried and was influenced by some of the work Horthy is doing at Humanlayer: "I tried every coding tool, and I experimented with basically every model. I just think Anthropic's models are the best at tool use right now, and that translates to code." Building a robust research engine The product itself mirrors this multi-agent approach. Stettner built Ambral's core research engine using the Claude Agent SDK with dedicated sub-agents for each data type. "I spent a lot of time using the Claude Agent SDK to basically build a very robust research engine that can operate across all of this data," Stettner explained. "It's based around Claude sub-agents, and for every type of data we have a dedicated sub-agent which is an expert in understanding that data." Whether users chat with the system or Ambral builds automations for customers, everything is backed by the Claude Agent SDK and a series of sub-agents retrieving and reasoning across usage data, Slack messages, meeting transcripts, and product interactions. The architectural inspiration came directly from Stettner's development experience: "I think how well Claude Code subagents were doing and helping me do development is what inspired me to basically want to take those same sub-agents and use it for the research engine in the product itself." Vulcan Technologies: Empowering non-technical founders to launch products For Tanner Jones, CEO and co-founder of Vulcan, Claude Code's impact extends far beyond productivity—it constitutes the democratization of company building. Founding their startup, the Vulcan team believed there had to be a product that could make government work better for citizens. That vision would have remained impossible without Claude Code because neither founder had an engineering background. Shipping a product without dedicated engineers Vulcan tackles a problem that's been accumulating for centuries: regulatory code complexity. Virginia's House of Burgesses, the oldest continuous democratic institution in the world, exemplifies this challenge. Regulatory buildup over 400+ years has created one of the most nuanced and complex codes in the U.S. When Aleksander Mekhanik and Tanner Jones co-founded Vulcan in April 2025, neither had a traditional engineering background. Mekhanik studied ML and mathematics in college, and Jones' last programming experience was an AP JavaScript class in high school where they wrote code with pen and paper. Yet the duo built a prototype of their first product for Virginia's governor's office by May 1st—and won the contract over established consulting firms. "The entire prototype was made using Claude," Jones explained. "This was pre-Claude Code. It was literally copy-pasting scripts into the web app, swapping out methods." After building the prototype, they hired their CTO, Christopher Minge, who had experience working at Google on Gemini and Waymo. Then, when Claude Code launched in June, the trio’s velocity multiplied again. Vulcan's AI-powered regulatory analysis helped reduce the average price of a new home in Virginia by $24,000, saving Virginians over a billion dollars annually by identifying redundant and duplicative regulatory requirements. Virginia’s governor loved Vulcan’s work so much that he signed Executive Order 51 , mandating that all state agencies use “agentic AI regulatory review.” Democratizing company building For Jones, Claude Code's impact goes beyond productivity metrics. "If you understand language and you understand critical thinking, you can use Claude Code well," he said. "I actually think there might be some marginal benefit for people who studied humanities, because the medium by which we're communicating with AI is language. If you have a great command of language and are good at constructing well-organized ordinal lists, nested bullet points and well-thought-out processes, your prompts may execute better." Jones commends Claude Code as a major component of Vulcan’s success: “In four months, with three founders, only one of whom was properly technical, we secured state and federal government contracts and raised an $11m seed round from some of the top VCs. None of this would have been possible without Anthropic’s unbelievable tools.” Christopher Minge, Vulcan’s CTO with “properly technical” training, experienced his own shift in how he thinks about engineering. "It feels a little bit like I have a co-worker at Google who I'm giving all of my ideas and tasks to, and they make mistakes frequently, but my role is delegating to several Claude Code instances and getting good at checking for common mistakes and communicating ideas effectively," Minge explained. Best practices from YC founders These three startups have developed battle-tested approaches to maximizing Claude Code's impact, including: 1. Separate research, planning, and implementation into discrete sessions "Don't make Claude do research while it's trying to plan, while it's trying to implement," Stettner advised. "Use discrete prompts and make those into discrete steps." This pattern prevents context contamination and allows each phase to focus on its core objective. Start a new Claude Code session for each major phase, passing only the distilled conclusions forward rather than dragging the entire context history. 2. Be deliberate about context management Stettner's advice for other founders centers on deliberate context management: "Context is critical. When I've seen output that was unexpected or low quality, it's generally due to a contradiction that I have in a prompt somewhere," he explained. "Be very deliberate in terms of what information you're putting into a system prompt or when you choose to start a new conversation, because you don't want to cloud your context. If there's any contradictions in your prompt, you're going to receive lower quality output." 3. Monitor and interrupt the chain of thought "Try to scrutinize the chain of thought and watch what it's doing," Jones suggested. "Have your finger on the trigger to escape and interrupt any bad behavior." This becomes especially important when running multiple instances. Catching a wrong direction early—within the first few tool calls—saves significantly more time than letting Claude Code complete an entire misguided approach. The new builder advantage These three startups demonstrate a fundamental shift in how companies are built with tools like Claude Code. HumanLayer pivoted and scaled while codifying context engineering practices that are now used across the YC ecosystem. Ambral is tackling customer success at massive scale with a lean founding team. Vulcan won government contracts as non-engineers. Traditional barriers to building software—technical expertise, team size, development time—are giving way to new competitive advantages: clear thinking, structured problem decomposition, and the ability to effectively collaborate with AI. Ready to build with Claude Code? Get started.

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Building multi-agent systems: When and how to use them¶

来源: Claude Blog 发布日期: Jan 23, 2026 采集时间: 2026-05-18 价值评分: 9.5/10 正文字数: ~21615 字符

摘要¶

Most teams don't need multi-agent systems. Learn the three scenarios where they consistently outperform single agents—and how to implement them effectively.

正文内容¶

While single-agent systems handle most enterprise workflows effectively, multi-agent architectures can unlock additional value for your organization. Learn when and how to use them.

A multi-agent system is an architecture where multiple LLM instances run with separate conversation contexts, coordinated through code. Multiple coordination patterns exist (agent swarms, capability-based systems, and message bus architectures), but this article focuses on the orchestrator-subagent pattern: a hierarchical model where a lead agent spawns and manages specialized subagents for specific subtasks. This pattern offers a straightforward coordination model and is a good starting point for teams new to multi-agent systems. We'll explore other patterns in detail in our next article. Today, multi-agent systems are often applied in situations where a single agent would perform better, though this calculus continues to evolve as models improve. At Anthropic, we’ve seen teams invest months building elaborate multi-agent architectures only to discover that improved prompting on a single agent achieved equivalent results. After building multi-agent systems and working with teams deploying them in production, we've identified three situations where multiple agents consistently outperform a single agent: when context pollution degrades performance, when tasks can run in parallel, and when specialization improves tool selection or task focus. Outside these situations, the coordination costs typically exceed the benefits. In this article, we share how to recognize single-agent limits, identify the three scenarios where multi-agent systems excel, and avoid common implementation mistakes. The case for starting with a single agent A well-designed single agent with appropriate tools can accomplish far more than many developers expect. Multi-agent systems introduce overhead. Every additional agent represents another potential point of failure, another set of prompts to maintain, and another source of unexpected behavior. We've observed teams build elaborate multi-agent systems with separate agents for planning, execution, review, and iteration, only to discover that they suffered from lost context at each handoff and spent more tokens coordinating than executing. In our testing, multi-agent implementations typically use 3-10x more tokens than single-agent approaches for equivalent tasks. This overhead stems from duplicating context across agents, coordination messages between agents, and summarizing results for handoffs. A decision framework for multi-agent systems Multi-agent architectures provide value when they address specific constraints that a single agent cannot overcome. This means multi-agent architectures should be reserved for cases where they provide clear benefits that justify the additional cost. The patterns below represent cases where we consistently observe positive returns on this investment. Context protection Large language models have finite context windows, and response quality can degrade as context grows. When an agent's context accumulates information from one subtask that is irrelevant to subsequent subtasks, context pollution occurs. Subagents provide isolation, with each operating in its own clean context focused on its specific task. Consider a customer support agent that needs to retrieve order history while diagnosing technical issues. If every order lookup adds thousands of tokens to the context, the agent's ability to reason about the technical problem degrades. ‍ The single-agent approach: # Single agent accumulates everything in context conversation_history = [ { "role" : "user" , "content" : "My order #12345 isn't working" }, { "role" : "assistant" , "content" : "Let me check your order..." }, # Tool result adds 2000 + tokens of order history { "role" : "user" , "content" : "... (order details, past purchases, shipping info) ..." }, { "role" : "assistant" , "content" : "Now let me diagnose the technical issue..." }, # Context is now polluted with order details the agent doesn 't need ] The agent must reason about the technical issue while maintaining 2000+ tokens of irrelevant order history in context, diluting attention and reducing response quality. ‍ The multi-agent approach: from anthropic import Anthropic client = Anthropic() class OrderLookupAgent : def lookup_order ( self , order_id : str ) -> dict : # Separate agent with its own context messages = [ { "role" : "user" , "content" : f "Get essential details for order {order_id}" } ] response = client.messages.create( model= "claude-sonnet-4-5" , max_tokens= 1024 , messages=messages, tools=[get_order_details_tool] ) # Returns only essential information return extract_summary(response) class SupportAgent : def handle_issue ( self , user_message : str ): if needs_order_info ( user_message ): order_id = extract_order_id(user_message) # Get only what 's needed, not full history order_summary = OrderLookupAgent().lookup_order(order_id) # Inject compact summary, not full context context = f"Order {order_id}: {order_summary[' status ']}, purchased {order_summary[' date ']}" # Main agent context stays clean messages = [ {"role": "user", "content": f"{context}\n\nUser issue: {user_message}"} ] response = client.messages.create( model="claude-sonnet-4-5", max_tokens=2048, messages=messages ) return response The order lookup agent processes the full order history and extracts a summary. The main agent receives only the 50-100 tokens it actually needs, keeping context focused. Context isolation is most effective when subtasks generate high context volume (more than 1000 tokens) but most of that information is irrelevant to the main task, when the subtask is well-defined with clear criteria for what information to extract, and for lookup or retrieval operations that require filtering before use. Parallelization Running multiple agents in parallel allows you to explore a larger search space than a single agent can cover. This pattern has proven particularly valuable for search and research tasks. Our Research feature uses this approach. A lead agent analyzes a query and spawns multiple subagents to investigate different facets in parallel. Each subagent searches independently, then returns distilled findings. Multi-agent search has shown substantial accuracy improvements over single-agent approaches by allowing exploration across larger information spaces. The core implementation decomposes a question into independent facets, runs subagents concurrently, then synthesizes the results. import asyncio from anthropic import AsyncAnthropic client = AsyncAnthropic() async def research_topic(query: str) -> dict: # Lead agent breaks query into research facets facets = await lead_agent.decompose_query(query) # Spawn subagents to research each facet in parallel tasks = [ research_subagent(facet) for facet in facets ] results = await asyncio.gather(*tasks) # Lead agent synthesizes findings return await lead_agent.synthesize(results) async def research_subagent(facet: str) -> dict: "" "Each subagent has its own context window" "" messages = [ { "role" : "user" , "content" : f "Research: {facet}" } ] response = await client.messages.create( model= "claude-sonnet-4-5" , max_tokens= 4096 , messages=messages, tools=[web_search, read_document] ) return extract_findings(response) This improved coverage comes at a cost. Multi-agent systems typically consume 3 to 10 times more tokens than single-agent approaches for equivalent tasks. This happens because each agent needs its own context, agents must exchange messages to coordinate, and results must be summarized when passed between agents. While parallelism helps reduce total execution time compared to running all that work sequentially, multi-agent systems often take longer overall than single-agent systems because of the sheer increase in total computation. The primary benefit of parallelization is thoroughness, not speed. When you need to search across a large information space or investigate many angles of a complex question, parallel agents can cover more ground than a single agent working within its context limits. The tradeoff is higher token usage and often longer total execution time in exchange for more comprehensive results. Specialization Different tasks sometimes benefit from different tool sets, system prompts, or domains of expertise. Rather than providing a single agent with access to dozens of tools, specialized agents with focused toolsets matched to their responsibilities can improve reliability. Tool set specialization When an agent has access to too many tools, performance suffers. Three signals indicate tool specialization would help: Quantity. An agent with too many tools (often 20+) struggles to select the appropriate one. Domain confusion. When tools span multiple unrelated domains (database operations, API calls, file system operations), the agent confuses which domain applies to a given task. Degraded performance. Adding new tools degrades performance on existing tasks, suggesting the agent has reached its capacity for tool management. System prompt specialization Different tasks sometimes require different personas, constraints, or instructions that conflict when combined. A customer support agent needs to be empathetic and patient; a code review agent needs to be precise and critical. A compliance-checking agent needs rigid rule-following; a brainstorming agent needs creative flexibility. When a single agent must switch between conflicting behavioral modes, separating into specialized agents with tailored system prompts produces more consistent results. Domain expertise specialization Some tasks benefit from deep domain context that would overwhelm a generalist agent. A legal analysis agent might need extensive context about case law and regulatory frameworks. A medical research agent might need specialized knowledge about clinical trial methodology. Rather than loading all domain context into a single agent, specialized agents can carry focused expertise relevant to their specific responsibilities. Example: Multi-platform integration. Consider an integration system where agents need to work across CRM, marketing automation, and messaging platforms. Each platform has 10-15 relevant API endpoints. A single agent with 40+ tools often struggles to select correctly, confusing similar operations across platforms. Splitting into specialized agents with focused toolsets and tailored prompts resolves selection errors. from anthropic import Anthropic client = Anthropic() # Specialized agents with focused toolsets and tailored prompts class CRMAgent : """ Handles customer relationship management operations """ system_prompt = "" "You are a CRM specialist. You manage contacts, opportunities, and account records. Always verify record ownership before updates and maintain data integrity across related records." "" tools = [ crm_get_contacts, crm_create_opportunity, # 8 - 10 CRM-specific tools ] class MarketingAgent : """ Handles marketing automation operations """ system_prompt = "" "You are a marketing automation specialist. You manage campaigns, lead scoring, and email sequences. Prioritize data hygiene and respect contact preferences." "" tools = [ marketing_get_campaigns, marketing_create_lead, # 8 - 10 marketing-specific tools ] class OrchestratorAgent : """ Routes requests to specialized agents """ def execute ( self , user_request : str ): response = client.messages.create( model= "claude-sonnet-4-5" , max_tokens= 1024 , system= "" "You coordinate platform integrations. Route requests to the appropriate specialist: - CRM: Contact records, opportunities, accounts, sales pipeline - Marketing: Campaigns, lead nurturing, email sequences, scoring - Messaging: Notifications, alerts, team communication" "" , messages=[ { "role" : "user" , "content" : user_request} ], tools=[delegate_to_crm, delegate_to_marketing, delegate_to_messaging] ) return response This pattern mirrors effective professional collaboration, where specialists with tools matched to their roles collaborate more effectively than generalists attempting to maintain expertise across all domains. However, specialization introduces routing complexity. The orchestrator must correctly classify requests and delegate to the right agent, and misrouting leads to poor results. Maintaining multiple specialized agents also increases prompt maintenance overhead. Specialization works best when domains are clearly separable and routing decisions are unambiguous. Outgrowing single-agent architectures Beyond the general framework, certain concrete signals suggest that single-agent patterns have been outgrown: Approaching context limits. If an agent routinely uses large amounts of context and performance is degrading, context pressure may be the bottleneck. Note that recent advances in context management ( such as compaction ) are reducing this limitation, allowing single agents to maintain effective memory across much longer horizons. Managing many tools. When an agent has 15-20+ tools, the model spends significant context and attention understanding its options. Before adopting a multi-agent architecture, consider using the Tool Search Tool , which lets Claude dynamically discover tools on-demand rather than loading all definitions upfront. This can reduce token usage by up to 85% while improving tool selection accuracy. Parallelizable subtasks. When tasks naturally decompose into independent pieces (research across multiple sources, tests for multiple components), parallel subagents can provide substantial speedups. These thresholds will shift as models improve. Current limits represent practical guidelines, not fundamental constraints. Context-centric decomposition When adopting a multi-agent architecture, the most important design decision is how to divide work between agents. We've observed that teams frequently make this choice incorrectly, leading to coordination overhead that negates the benefits of multi-agent design. The key insight is to adopt a context-centric view rather than a problem-centric view when decomposing work. Problem-centric decomposition (often counterproductive). Dividing by type of work (one agent writes features, another writes tests, a third reviews code) creates constant coordination overhead. Each handoff loses context. The test-writing agent lacks knowledge of why certain implementation decisions were made and the code reviewer lacks the context of exploration and iteration. Context-centric decomposition (usually effective). Dividing by context boundaries means an agent handling a feature should also handle its tests, because it already possesses the necessary context. Work should only be split when context can be truly isolated. This principle emerges from observing failure modes in multi-agent systems. When agents are split by problem type, they engage in a "telephone game," passing information back and forth with each handoff degradi

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

How Carta Healthcare gets AI to reason like a clinical abstractor¶

来源: Claude Blog 发布日期: Apr 08, 2026 采集时间: 2026-05-18 价值评分: 9.5/10 正文字数: ~5253 字符

摘要¶

How Carta Healthcare used Claude and context engineering to build Lighthouse, a clinical abstraction platform reaching 99% accuracy across 22,000 cases a year.

正文内容¶

Inside a clinical data abstraction platform processing 22,000 surgical cases a year—and how the team reached 99% accuracy by getting context engineering right. ‍

In our new series, How startups build with Claude , we highlight how hypergrowth organizations are transforming their industries with AI. In this article, we share engineering lessons learned behind the creation of Lighthouse, Carta Healthcare's clinical data abstraction platform, and why context engineering matters as much as model capability when you’re building AI-powered systems at scale. The quick pitch Name Carta Healthcare Founded 2017 CEO Brent Dover Stack Claude in Amazon Bedrock Growth 10x growth in the last 3 years, supporting 125+ hospitals

Accelerate your enterprise AI transformation with proven strategies from Anthropic's customers in HCLS.

Building the right context Every data point Lighthouse extracts needs different source documents, a different time window, and a different amount of context to answer correctly. This is the core challenge of context engineering: an AI agent's performance isn't determined solely by the model. It's determined by what the model is given to work with. The best prompt in the world underperforms if it's fed incomplete, unordered, or poorly scoped information. Getting context right—what to include, what to exclude, and in what sequence—is the engineering work that separates a prototype from a production system. Take pre-procedure weight. The requirement isn't "find a weight in the record." It's "find a weight documented before a specific procedure start time." The most recent weight may have been recorded two days after the operation. That value is inadmissible. Carta Healthcare’s system incorporates patient-specific context at runtime so Claude has a precise time boundary for each question. The prompt for pre-procedure glucose specifies the procedure start time and asks for the most recent glucose before that moment. Without that specificity, even a strong model is working from incomplete information that any trained abstractor would catch immediately. “The hardest problems we solved weren't about building a perfect prompt, they were about context construction,” says Matthew Mazzanti, Software Engineering Manager at Carta Healthcare. For Carta Healthcare’s team, the hard work was building the pipeline that assembles the right information at runtime—the right documentation, the right timeframe, the right priority order. “Integrating, organizing, and surfacing the right data at the right time is the real work. A perfectly written prompt with bad context gives bad answers. A straightforward prompt with the right context delivers the results you need," says Mazzanti. Glaser's advice for teams in the same position? Start by asking whether you're giving the model what it needs to reason, or asking it to figure things out from whatever's available. "When something underperforms, you can trace it back to a specific prompt, a context issue, or a retrieval gap rather than staring at an aggregate score wondering what went wrong," says Mazzanti. "Build your evaluation framework early, make it granular, and design it to isolate variables. Skip this, and you'll spend more time debugging than building." At one large health system, Lighthouse processed over 22,000 surgical cases annually across 14 hospitals, with inter-rater reliability reaching 98-99% , the industry's standard measure of abstraction accuracy. Keeping clinical expertise in the loop Once context construction is solid, the people who understand clinical documentation best can start shaping how the system behaves. Building trust with abstractors came down to transparency. Lighthouse isn't a black box. For every data point it extracts, abstractors see the supporting evidence and Claude's rationale. They can validate findings and exercise clinical judgment rather than accepting outputs. From there, prompting becomes the mechanism through which clinical expertise directly shapes how Claude reasons. When an abstractor finds that a specific registry data point isn't being extracted correctly, her explanation of why—the edge cases, the documentation patterns, what the prompt is missing—becomes a direct input to how Claude handles that field. By prioritizing context engineering early on, Carta Healthcare turns that explanation into a revised prompt and ships it the same day. "Our clinical abstractors regularly hand us long explanations of how a specific data point works in practice," Glaser says. "Instead of spending weeks translating that into data science models and custom code, we use that feedback directly in the prompts. What used to take months of engineering and QA per registry now ships in a week." As one of Carta Healthcare’s abstractors put it: "Lighthouse doesn't replace my judgment. It enhances it." To learn more, read Carta Healthcare's full story . ‍

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Claude Code and new admin controls for business plans¶

来源: Claude Blog 发布日期: Aug 20, 2025 采集时间: 2026-05-18 价值评分: 9.5/10 正文字数: ~5021 字符

摘要¶

Enterprise and Team customers can now access Claude Code through premium seats, with new admin controls, usage analytics, and a Compliance API for governance.

正文内容¶

Claude brings its conversational AI and powerful coding agent together in one enterprise subscription.

Enterprise and Team customers can now upgrade to premium seats that include more usage and Claude Code—bringing our app and powerful coding agent together under one subscription. Users can move seamlessly between ideation and implementation, while admins get the visibility and controls they need to scale Claude across their organization. We are also introducing a new Compliance API, giving organizations programmatic access to usage data and customer content for better observability, auditing, and governance. Add Claude Code to your plan Admins have full flexibility to assign standard or premium seats according to individual user requirements and organizational roles. Premium seats give users access to both Claude and Claude Code, enabling them to partner with Claude throughout the entire development lifecycle. For example, developers can research unfamiliar frameworks through interactive chats with Claude, then implement production-ready code using Claude Code in their terminal. They can explore architectural approaches and evaluate trade-offs in Claude conversations, then switch to Claude Code to build their chosen architecture. Early customers like Behavox, a compliance and security company, have seen immediate results. “Since bundling Claude for Enterprise with Claude Code, we’ve rolled it out to hundreds of developers, and it has quickly become our go-to pair programmer,” shares Artem Pikulin, Senior Manager of Machine Learning Operations. “Its coding assistance consistently outperforms other agents, delivering superior results and value every day.” Altana, the AI-powered Product Network connecting the global supply chain, has experienced similar transformational results. “Claude Code and Claude have accelerated Altana's development velocity by 2-10x, transforming how we build sophisticated AI/ML systems that facilitate multi-party collaboration on large-scale knowledge graphs of global supply chains,” says Peter Swartz, Co-founder and Chief Science Officer. “We're taking on significantly more ambitious projects as a result.” Flexible pricing, predictable costs Claude seats include enough usage for a typical workday, but for times when your teams need access to more intelligence and additional conversations with Claude–admins can enable extra usage for individual users at standard API rates. Admins have control over the maximum amount a user can spend with extra usage to ensure that users get flexibility and admins get predictable billing. This approach provides a simple way to scale with Claude, while offering centralized billing, management, and enterprise-grade administrative controls at every step. Admin control and management We've built new comprehensive controls that give organizations the visibility and management capabilities they need to enable employees to work productively with Claude. Self-serve seat management : Purchase new seats, directly manage seat allocation, and provision users through the admin panel. Granular spend controls: Set spending limits at the organization and individual user level to stay within budget while maintaining flexibility for essential projects. Usage analytics : View Claude Code analytics in Claude, including metrics like lines of code accepted, suggestion accept rate, and usage patterns. Managed policy settings : Deploy and enforce settings across all Claude Code users to match internal policies, including tool permissions, file access restrictions, and MCP server configurations. Compliance API Enterprise organizations can now better meet regulatory requirements with our new Compliance API. Rather than manual exports and periodic reviews, compliance teams get real-time programmatic access to Claude usage data and customer content, enabling them to build continuous monitoring and automated policy enforcement systems. Administrators can integrate Claude data into existing compliance dashboards, automatically flag potential issues, and manage data retention through selective deletion capabilities. This provides the visibility and control organizations need to scale AI adoption while meeting regulatory obligations. Getting started Team and Enterprise plan admins can now upgrade to premium seats with Claude Code—and take advantage of flexible pricing with extra usage options. Both Team and Enterprise plans include granular spend caps, self-serve seat management, and Claude Code usage analytics. To explore how Claude for Enterprise can transform your organization's productivity and learn more about the Compliance API, contact our sales team.

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Long context prompting for Claude 2.1¶

来源: Claude Blog 发布日期: Dec 06, 2023 采集时间: 2026-05-18 价值评分: 9.0/10 正文字数: ~494 字符

摘要¶

Claude 2.1 excels at retrieving information across its 200K context window, with a simple prompt adjustment improving accuracy from 27% to 98%. Claude 2.1’s performance when retrieving an individual sentence across its full 200K token context window.

正文内容¶

Claude 2.1 excels at retrieving information across its 200K context window, with a simple prompt adjustment improving accuracy from 27% to 98%.

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Claude 2 on Amazon Bedrock¶

来源: Claude Blog 发布日期: Aug 23, 2023 采集时间: 2026-05-18 价值评分: 9.0/10 正文字数: ~3659 字符

摘要¶

Claude 2 is now available on Amazon Bedrock, helping enterprises like LexisNexis, Lonely Planet, and Ricoh USA build secure, scalable AI applications. We’re excited that Claude 2 is now available to customers on Amazon Bedrock.

正文内容¶

Claude 2 is now available on Amazon Bedrock, helping enterprises like LexisNexis, Lonely Planet, and Ricoh USA build secure, scalable AI applications.

We’re excited that Claude 2 is now available to customers on Amazon Bedrock. Claude 2 is our state-of-the-art model that excels at thoughtful dialogue, content creation, complex reasoning, creativity, and coding, built with Constitutional AI . Amazon Bedrock is a fully managed service that makes leading foundation models accessible via an API so it’s easier for businesses to build and scale generative AI-based applications. When the service launched in April of this year, Claude 1.3 and Claude Instant were among the first foundation models available to Amazon Web Services (AWS) customers. Here are a few examples of teams building with the newly available Claude 2 on Amazon Bedrock: LexisNexis and AI-assisted legal services LexisNexis Legal & Professional is a leading global provider of information and analytics serving customers in more than 150 countries across law firms, corporate, government, and academic markets, and is committed to responsibly developing AI solutions with human oversight. Claude 2 via Amazon Bedrock will help power LexisNexis’ legal AI capabilities. “It’s important to understand AI’s capabilities and limitations, and to work with partners that deliver data security and provide best-in-class training of large language models. Legal use cases also require high-quality technical analysis, long context windows for processing detailed documents, and fast outputs. That’s why we’ve chosen Claude 2 on Amazon Bedrock as an important part of our AI strategy,” said Jeff Reihl, Executive Vice President & Chief Technology Officer at LexisNexis Legal & Professional. Lonely Planet and AI-supported travel planning Lonely Planet is one of the world’s most trusted sources of travel information. Its team is using Claude 2 on Amazon Bedrock to unlock decades of travel content and help customers plan memorable trips. “We’ve turned to Claude 2 on Amazon Bedrock to integrate generative AI in a scalable, reliable, and secure way, making it easier than ever for our customers to access our world-class travel content when and how they want,” said Chris Whyde, Senior Vice President of Engineering and Data Science at Lonely Planet. Ricoh USA and AI-driven operations Ricoh USA offers workplace solutions and digital transformation services designed to manage and optimize the flow of information across businesses. It’s using Claude 2 on Amazon Bedrock to infuse generative AI into its operations. “We wanted to integrate a large language model that could produce high-quality training datasets while preserving data integrity and customer security. We selected Claude 2 on Amazon Bedrock for its proficiency in generating this data, as well as its adherence to stringent standards including HIPAA and SOC II,” said Ashok Shenoy, Vice President of Portfolio Development at Ricoh USA. Getting started To learn more about building next-generation AI applications with Claude on Amazon Bedrock, please visit aws.amazon.com/bedrock/claude . We’re grateful for the interest in this important collaboration, and we look forward to making our safer, more reliable AI models even more accessible through Amazon Bedrock in the future.

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Claude and Slack¶

来源: Claude Blog 发布日期: Oct 01, 2025 采集时间: 2026-05-18 价值评分: 9.0/10 正文字数: ~1229 字符

摘要¶

Access Claude directly into Slack so you can get help in channels and threads, or connect your Slack workspace to Claude for searching conversations and pulling context when you need it.

正文内容¶

Access Claude directly into Slack so you can get help in channels and threads, or connect your Slack workspace to Claude for searching conversations and pulling context when you need it.

Update: The Slack connector is now available for Claude Pro and Max subscribers. (January 26, 2026) We're introducing two ways to use Claude with Slack. Add Claude directly to your Slack workspace, or connect Slack to the Claude apps so Claude can search and reference relevant Slack messages during conversations. With Claude and Slack deeply integrated, you can work wherever you’d like. Get Claude's help directly in Slack conversations, or have Claude search your Slack workspace when you need context for deeper research projects. Draft responses to critical messages, prep for meetings by pulling in relevant discussions, or analyze shared documents—all where you are already working.

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Claude API skill now in CodeRabbit, JetBrains, Resolve AI, and Warp¶

来源: Claude Blog 发布日期: Apr 29, 2026 采集时间: 2026-05-18 价值评分: 9.0/10 正文字数: ~2352 字符

摘要¶

Today, CodeRabbit, JetBrains, Resolve AI, and Warp are bundling the claude-api skill, giving developers production-ready Claude API code wherever they build. First introduced in Claude Code in March, the skill is now in more of the tools developers already use.

正文内容¶

Today, CodeRabbit, JetBrains, Resolve AI, and Warp are bundling the claude-api skill , giving developers production-ready Claude API code wherever they build. First introduced in Claude Code in March, the skill is now in more of the tools developers already use. Building with the Claude API skill The claude-api skill captures the details that make Claude API code work well, like which agent pattern fits a given job, what parameters change between model generations, and when to apply prompt caching. The result is fewer errors, better caching, cleaner agent patterns, and smoother model migrations. It stays current as our SDKs change. When a new model is released or the API gains a feature, Claude already knows. Anywhere the skill is available, ask Claude to: "Improve my cache hit rate." The skill applies prompt caching rules many developers miss. "Add context compaction to my agent." It walks you through the compaction primitives and agent patterns in our docs. "Upgrade me to the latest Claude model." Claude reviews your code and walks you through updating model names, prompts, and effort settings for a new model like Opus 4.7 . In Claude Code, you can also run this directly with /claude-api migrate. ‍ "Build a deep research agent for my industry." Claude walks you through configuring Claude Managed Agents , so long-running research is a few prompts, not a custom project. In Claude Code, you can also run this directly with /claude-api managed-agents-onboard .

For Claude-powered coding agents Any coding agent can bundle the claude-api skill to give their users expertise around the Claude API. If you are building a tool where developers write Claude API code, the skill is open source at anthropics/skills . Our bundling guide walks through the setup in about 20 lines of CI, and the skill stays current automatically. Getting started The skill is already in Claude Code , CodeRabbit , JetBrains , Junie , Resolve AI , and Warp . To learn more, see the claude-api skill docs .

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集

Claude now creates interactive charts, diagrams and visualizations¶

来源: Claude Blog 发布日期: Mar 12, 2026 采集时间: 2026-05-18 价值评分: 8.2/10 正文字数: ~790 字符

摘要¶

Ask Claude to explain a concept or analyze your data, and it can respond with interactive charts, diagrams, and visualizations — rendered inline as part of the conversation.

正文内容¶

Update: Now available in Claude Cowork on all paid plans (April 22, 2026). Last fall, we previewed Imagine with Claude : a new way for Claude to build visuals in real time, without any code. We’re now bringing a version of this feature, in beta, to Claude’s chat conversations. Claude can create custom charts, diagrams and other visualizations in-line in its responses—and then tweak and modify its creations as the conversation develops.

Explore more product news and best practices for teams building with Claude.

Transform how your organization operates with Claude

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.

采集自 Claude Blog，由 collect_claude_blog.py 自动采集