Anthropic · 2026-05-06 · major
Claude Managed Agents — Multi-Agent Sessions and Outcomes Hit Public Beta
Anthropic moves multi-agent orchestration and outcome-driven sessions to public beta for Claude Managed Agents. A coordinator delegates to up to 20 specialist agents; outcome rubrics let agents self-evaluate and iterate.
Coordinator agents, specialist roles, and rubric-graded iteration land as a single API surface for production agent stacks.
Key specs
| Concurrent threads | 25 max |
|---|---|
| Specialist agents per coordinator | 20 max |
| Outcome iterations | default 3, max 20 |
| Beta header | managed-agents-2026-04-01 |
What is it?
Two new public-beta features for Claude Managed Agents — Anthropic's hosted agent harness — landed on May 6, alongside the Code w/ Claude 2026 keynote in San Francisco. Multi-agent sessions let a coordinator agent delegate to a roster of up to 20 specialist agents, each with its own model, system prompt, tools, and isolated context. Outcomes let you attach a markdown rubric to a session and have a separate grader judge each iteration until the rubric is satisfied.
How does it work?
All agents share one container and filesystem but run in separate session threads with independent conversation history. The coordinator emits cross-thread events on a primary stream and can spawn up to 25 concurrent threads. For outcomes, a user.define_outcome event attaches a rubric (inline text or via the Files API) and a max_iterations cap that defaults to 3 and tops out at 20. A grader runs in a separate context window and returns per-criterion pass or fail; failed criteria feed back to the agent for the next loop. Both features ship under the existing managed-agents-2026-04-01 beta header. The same release also adds webhooks for session and vault lifecycle events plus background refresh for mcp_oauth vault credentials.
Why does it matter?
These were the missing primitives for production agent stacks on Claude. Without coordinator and specialist split, teams had to bolt their own orchestration around a single agent. Without outcomes, the only stop signal was the model's own judgment, which doesn't survive long-horizon work. Together they turn Managed Agents from a single-agent harness into a configurable platform with built-in self-evaluation, putting Anthropic on a more direct collision course with Mistral Workflows and Google's Gemini Enterprise Agent Platform.
Who is it for?
Teams building autonomous agent products on the Anthropic API
Try it
anthropic-beta: managed-agents-2026-04-01