If you're perfectly qualified to do something, you've already outgrown it

You can now discover tools, agents and expert partners on Claude Marketplace. Use it to: - Add connectors and plugins like Slack and Notion - Buy agents and products from companies like Cursor and CrowdStrike - Scale with service partners like Accenture and Deloitte

This was done by a fleet of 950 agents working 21 hours and using 210 million tokens. Long running agentic tasks can unlock new discoveries.

We had early access to Opus 5.5 and ran it on Ramp Accounting Bench. It performs close to Fable 5.1, at 61% lower cost and 1.7x faster.

We just shipped Opus 5.5. It's top tier on the benchmarks, 40% cheaper, and 30% faster than its predecessor. Plus talking to it is easy again :)

Live now: our Harness Engineering Track from AI Engineer World's Fair 2026. Ports, proofs, kill switches, and what is left of an agent when you remove the model. Thesis: when an agent breaks in production, the model is rarely the part that failed. piped.video/watch?v=PXj0p_mW… - 2026 State of AI Engineering: @barrnanas, Amplify Partners - The Unreasonable Effectiveness of Separating the Task from the Model: @MaximeRivest & @isaacbmiller1, DSPy - How Anthropic Builds: @mikeyk, Anthropic - Tokens Should Have Jobs: @katelyn_lesse & @angjiang, Anthropic - Building the Production Cage for Powerful Domain Agents: Mike Chambers, AWS - Loophole: @brendanh0gan, Morgan Stanley - Every step you take, every call you make: @vdgiselle, Restate - Agent Frameworks Considered Harmful: @remilouf, .txt - We let an AI agent execute Bash and lived to talk about it: Sarah Sanders, PostHog - No Memory, No Harness: Kay Malcolm, Oracle - How we Solved Agent Building: @andrewqu, Vercel - Agents Without Code: @_philschmid, Google DeepMind - "I've never seen anything scarier than an LLM with tool calls." @headinthebox

The importance of a task is measured by the consequence of its performance in either positive or negative extremes. If there are no consequences in either direction, the task is not useful.

The evolution of human-ai interaction Human: ask -> delegate -> entrust AI: respond -> execute -> anticipate

Point Claude to this repo. Tell it to build your own. Profit.

For the Season 2 premiere of Builders, @josh_coyne and I sat down with @katelyn_lesse and @angjiang from Anthropic to learn from their expertise building & using agents. We also dug into Anthropic's roadmap and what they plan to build vs. working with startups. Link below!

Fable gets a glow up. SOTA on the benchmarks, very excellent on low reasoning, and with a 75% drop on api cache reads for those long running agentic workloads

Today, we're kicking off the first phase of the research preview for Model Hardware Standard (MHS): a new standard for AI agents to safely operate physical equipment in scientific research and advanced manufacturing. Read more: anthropic.com/news/model-har…

“Harness” does not mean application. An application can have many harnesses. A harness is a loop that runs the model. Its purpose is to get the model to give you a good answer. The application uses the harness(es) to give you a good experience.

This is why it’s better to use the API directly from the model provider. Yes, harness matters. The API itself is the lowest level harness and different models have different expressions.

A lot of moats are moving from first order to second order. A first order moat tends to defend a product and is mostly about technology. A second order moat tends to surround the thing that makes the product and defends the rate of adaptation.

Who touches every token that flows through @anthropic? It’s not any one model, but it’s @katelyn_lesse, @angjiang and the platform team. A year ago, it was just a messages API. Today, their platform sits under every API call, every workflow, and every agent built on Claude. @sonyatweetybird and I explore why they make everything they can externally facing for customers, how they pick markets to target, what it takes to build the best agents, and more: 00:00 Introduction 01:49 Two North Stars 02:27 External Builders And Primitives 03:54 What To Externalize 06:00 From Messages To Agents 08:19 Managed Agents Adoption 09:07 Three Layer Cake 10:22 Execution Harnesses Explained 11:09 Coordination Strategies Roadmap 12:13 Ecosystem Standards And Safety 15:39 Open Ecosystem Not Walled 17:12 Vertical Products And Form Factors 22:26 Claude Tag Under The Hood 26:04 Harness Best Practices 38:13 Token Costs And Whats Next

The final AI api only has two parameters: outcome and budget. It has a signing for who made the request. Then everything is just built by the AI as needed. And what’s left are the human things: 1) what to do; 2) how much it’s worth; 3) and who is accountable

This model is SOTA on everything and available in two flavors: 1) Fable 5 for agentic systems 2) Mythos 5 for cyber and biomedical research acceleration

Meaningful economic output is driven by knowledge, execution, or coordination. Value creation comes from increasing the capability, reducing the cost, or improving the speed of one of these three functions.