Anthropic Blocks Third‑Party Agents, Then Launches Claude Managed Agents to Disrupt the Startup Scene

Anthropic has released Claude Managed Agents, a fully hosted agent deployment platform where users only need to define an agent’s task, tools, and rules and Anthropic runs it, providing sandboxed execution, state management, credential isolation and out‑of‑the‑box multi‑agent coordination.

"Anthropic just mass‑obsoleted every agent orchestration startup in a single launch." – Aakash Gupta

The author calls this a “dimensionality‑reduction strike” because existing frameworks such as LangChain, CrewAI and AutoGen act as glue on top of model‑provider APIs, whereas Anthropic now offers a native solution built by the model supplier itself.

Key capabilities include:

Safe sandbox execution : agents run in isolated environments with credentials stored outside the sandbox, preventing prompt‑injection attacks.

Long‑running sessions : agents can operate for hours without losing progress even if disconnected.

Multi‑agent coordination (research preview) : one agent can spawn and direct other agents to handle tasks in parallel.

MCP native integration : OAuth token management and Claude calls MCP tools via a dedicated proxy.

Session persistence : an immutable event stream enables context recovery and history replay.

These features are specifically tuned for Claude, whereas third‑party frameworks remain generic.

The architecture is split into three decoupled components:

Brain : the Claude model plus a Harness execution framework that decides and invokes tools via a single interface execute(name, input) → string.

Hands : the execution environment comprising sandbox containers, code executors, external tools and the MCP server; Brain and Hands communicate through the unified interface and are not bound to the same container.

Session : an immutable event stream independent of Harness that supports slice queries and context restoration, cleanly solving the “long‑dialog exceeds context window” problem of traditional agent frameworks.

Anthropic reports concrete performance gains: p50 first‑token latency drops about 60% and p95 latency improves by more than 90%, because the stateless Harness only allocates containers when needed instead of dedicating a container per agent.

Fault‑tolerance is built in: container failures are caught by Harness, which lets Claude decide whether to retry by launching a fresh container; Harness failures are recovered by calling wake(sessionId) to replay the session event stream and continue execution. Since Session data is persisted externally, the system remains resilient even if individual components crash.

Early customers listed by Anthropic include Notion (parallel task execution), Rakuten (enterprise‑wide agents across product, sales, marketing, finance and HR with Slack/Teams integration), Asana (an “AI teammate” that collaborates on tasks and drafts deliverables) and Sentry (debugging and patch‑writing agents that go from bug tagging to reviewable PRs in weeks). These are production deployments, not demos, and Anthropic claims a ten‑fold speedup from prototype to launch.

The launch follows a strategic move: four days earlier Anthropic blocked third‑party Harness access, then introduced Managed Agents as a superior, in‑house alternative. This mirrors a broader trend where model providers are internalizing agent infrastructure—OpenAI with Assistants API, Google with Vertex AI Agent Builder, and now Anthropic with Managed Agents—compressing the space for third‑party orchestration frameworks unless they focus on vertical depth or multi‑model orchestration value.