Over the past 45 days, Anthropic has hardly stopped.
Fable 5 and Mythos 5 are launching simultaneously, Sonnet 5 is price-reduced to gain market share, and Claude Code is receiving nearly daily updates.
Cowork was just updated yesterday, and today Anthropic has rolled out six updates at once on its Managed Agent Platform (CMA):
- Skill cap raised from 20 to 500: A single session can hold an entire enterprise knowledge base.
- Five-level intensity adjustment: Simple routing saves tokens, deep reasoning maximized
- One-step seed session creation: directly includes 50 initial events, zero cold start
- Sub-agents stream in real time down to the thread level: where each sub-agent goes and what it thinks is visible in real time.
- Seven new webhooks cover the entire lifecycle of environment and memory storage: polling can now be retired.
- Agent update version field made optional: eliminates version management overhead in single workflows
A single session can run 500 skills and power an entire enterprise knowledge base—Anthropic is essentially building an operating system for agents!
Skill cap increased to 500—capable of handling an entire business department in a single conversation.
The CMA skills are knowledge packages that inject professional instructions into the agent, previously limited to a maximum of 20.
However, a financial customer service agent alone must handle七八 tasks—account opening, KYC, anti-money laundering, complaint resolution, compliance disclaimers, and multilingual switching—plus regulatory differences and A/B test versions, immediately hitting a ceiling.
This number has now been directly raised to 500.
These 500 skill slots are shared by all agents within a single session, running in a unified hosted sandbox, with agents sharing memory and environment variables, coordinated and scheduled centrally—sufficient to represent a complete business unit.
Moreover, you can even include the entire codebase standards, CI pipeline configuration, and deployment targets in a single session, allowing the agent to understand everything from coding to deployment.
Five-level intensity adjustment—don’t exhaust your full token power on simple questions.
The effort parameter has long existed at the API level, but agents in CMA previously could not be set individually.
In other words, the entire session runs on the default setting, so asking “What’s today’s date?” and “Audit a 500-page financial report” use the same level of depth. Over-reasoning simple questions is purely wasting tokens.
The thinking intensity can now be directly set in each agent’s model configuration, with five levels: low / medium / high / xhigh / max.
You run a multi-agent customer service system: routing for coordinators uses low, product inquiries use medium, complaint escalations use high, and legal compliance uses max.
Within the same session, different agents can operate at varying performance levels based on task complexity. Finally, there is a continuous dial balancing cost and quality.
Cold start reset to zero, sub-agent black box also disassembled.
Before creating a CMA session, two steps are required: first, POST an empty shell, then sequentially inject initial events. In high-concurrency scenarios, round-trip latency accumulates layer by layer, significantly slowing down cold start performance.
Go straight for it now. POST /v1/sessions with initial_events, up to 50 initial events, so the session knows exactly what to do from the moment it’s created.
In the same update, the version field in the agent update API also became optional.
Previously, each update required submitting a version number for concurrency control, and a version mismatch would return a 409 error. Now, in single-process scenarios, this step can be skipped entirely.
The observability of sub-agents has also made significant progress.
At the end of June, CMA launched session-level event_deltas, enabling real-time previewing of the main agent's text output. However, sub-agents remain black boxes—whether they go off track or freeze, you can only discover it after a timeout occurs.
This time, we directly go down to the thread level. Subscribe to independent thread streams of sub-agents—write wherever you think, and see it in real time.
7 new webhook slots—polling is over.
CMA launched webhooks for agents and deployment lifecycles at the end of June.
This update completes the remaining two sections: four types of environmental events and three types of memory storage events.
At this point, the entire lifecycle of CMA—environment setup, agent deployment, runtime status, and memory management—can be fully integrated with external systems through event-driven mechanisms.
Automatically trigger alerts when the environment is paused, synchronize updates to your own database when memory storage is updated, and automatically attach monitoring when a new environment is created.
From "API products" to "an operable platform"—this is the final piece of the puzzle.
Anthropic is betting on OS-level depth.
By July 2026, the competition for AI agent infrastructure has entered the deep water stage, with update frequencies nearly daily.
OpenAI bets on an entry point: cramming chat, programming, and autonomous work into a single ChatGPT super-app to lock in developers with its massive user base.
Google bets on infrastructure: GCP + TPU + protocol standards, building a moat at the foundational level.
Anthropic bets on OS-level depth: hosted sandbox + full-stack agent control + adjustable reasoning intensity, focusing on increasing the capability density of a single agent.
From 20 to 500, Anthropic is calculating when agents will become as fundamental as an operating system.
The speed has already answered: very fast.
Reference: https://x.com/ClaudeDevs/status/2080009523952263295?s=20
This article is from the WeChat public account "New Intelligence Yuan," author: ASI Revelation; editor: Moses