Anthropic adds managed agents docs

Anthropic’s new managed-agents docs matter because they show what the company thinks “agent infrastructure” actually is. Not just a model that can call tools. A hosted runtime with sessions, containers, built-in tools, event history, and tracing. The gap until now was that “agents” often meant a pile of glue code developers had to write themselves. Anthropic’s docs make the pitch explicit: let Claude run inside Anthropic’s harness instead. ### What changed here? The news is less a splashy launch than a documentation line getting filled in. Anthropic already put Claude Managed Agents into public beta on April 8, 2026. But the docs now lay out the whole product surface in a way that makes it usable as a platform feature, not just a beta announcement — overview, quickstart, agent setup, sessions, tools, observability, and even multi-agent preview pages. (platform.claude.com) ### What is a managed agent, exactly? Basically, Anthropic is offering a prebuilt agent loop. Instead of wiring up model calls, tool dispatch, retries, state storage, and container execution yourself, you define an agent once and run it in Anthropic’s managed environment. The docs frame it as “fully managed agent infrastructure” where Claude can read files, run commands, browse the web, and execute code securely inside a hosted runtime. (platform.claude.com) ### How is the product structured? Anthropic breaks it into four pieces. An agent is the reusable configuration — model, system prompt, tools, MCP servers, and skills. An environment is the container template. A session is the running instance that combines the two and keeps history. And events are the stream of messages, tool results, and status updates flowing through the session. That split is the real produ(platform.claude.com)nd “what job it’s doing right now.” (platform.claude.com) ### Why does that matter to developers? Because the annoying part of agents is usually not the prompt. It’s the orchestration. Long-running work, async jobs, retries, file handling, sandboxing, and reconnecting to state after something pauses — that’s the stuff teams end up rebuilding. Anthropic is trying to turn that into infrastructure. Sessions are stateful. Responses can stream over server-sent events. Agents are versio(platform.claude.com 1) (platform.claude.com 2) ### What tools does Claude get? The managed runtime includes a built-in toolset, and developers choose which tools an agent can use. Anthropic also supports custom tools, where the app executes the tool and sends the result back, plus MCP servers and skills layered into the agent definition. In other words, Anthropic is not just selling “Claude with functions.” It’s selling a menu of capabilities inside a controlled runtime. (([platform.claude.com)# Where does observability show up? This is the part that makes the docs more serious than a demo. Anthropic exposes a session timeline in the Console with status, timestamps, token usage, and tool execution details. That means the company knows the trust problem with agents is not only “can it do the task?” but “can a human see what happened?” The cookbook examples push the same idea — an incident-response agent can draft a fix PR, then stop for approval before merging. (platform.claude.com) ### Is this just another agent wrapper? Not quite. The catch is that Anthropic is moving one layer down the stack. Plenty of companies offer agent frameworks. Managed Agents is closer to a hosted runtime plus control plane. Even the API tells that story — separate Agents, Sessions, and Environments endpoints, with a dedicated beta header dated `2026-04-01`. That looks more like infrastructure than a convenience feature. (platform.claude.com) ### What’s the bottom line? Anthropic is trying to make “run an agent” feel like “deploy a service.” That could save teams a lot of plumbing work. But the docs also reveal the real bar for these systems: not autonomy alone — visible traces, tool logs, and human approval points that make the autonomy safe to use. (platform.claude.com)