Claude Managed Agents
Claude Managed Agents is a suite of composable APIs from 📝Anthropic for building, running, and observing cloud-hosted AI agents at scale, launched in public beta on April 8, 2026. The platform decouples the model — the "brain" — from execution infrastructure — the "hands" — by providing a fully managed, sandboxed environment in which 📝Claude can read files, run commands, browse the web, and execute code under scoped credentials.
Key Facts
- Status: Public beta (requires
managed-agents-2026-04-01beta header; SDK sets it automatically) - Access: Enabled by default for all API accounts
- Pricing: Standard token pricing + $0.08/session-hour (runtime only, measured to the millisecond) + $10/1k web searches
- CLI:
antCLI available via Homebrew (macOS), curl (Linux), or Go install - SDKs: Python, TypeScript, Java, Go, C#, Ruby, PHP
- Rate limits: 60 RPM for create endpoints, 600 RPM for read endpoints (org-level)
- Batch API discounts do not apply — Managed Agents is billed separately from the standard Messages API
- Research preview features: Outcomes and memory (require access request); multiagent orchestration has since moved to beta (see Multiagent Orchestration below)
Core Concepts
The platform is built around four primitives:
How It Works
- Create an agent — define model (e.g.,
claude-opus-4-7), system prompt, and tools - Create an environment — configure a cloud container with networking rules
- Start a session — launch a session referencing your agent and environment
- Send events and stream responses — send user messages as events; Claude autonomously executes tools and streams results via SSE
- Steer or interrupt — send additional events to guide the agent mid-execution, or interrupt to change direction
When a session starts, the agent loop runs autonomously: Claude decides which tools to use, executes them inside the sandboxed container, and streams events back. The session emits session.status_idle when complete. Sessions persist file systems and conversation history across multiple interactions.
Built-in Tools
The agent_toolset_20260401 tool type enables all built-in tools:
Tools can be individually enabled/disabled via a configs array. Custom tools (client-executed, like Messages API tool use) are also supported alongside the built-in toolset.
Architecture: Brain-Hands Decoupling
The engineering design follows an operating systems principle: virtualizing components into stable abstractions that outlast implementations.
Session — a durable, append-only event log stored outside Claude's context window. Supports crash recovery via wake(sessionId). The session functions as "a context object that lives outside the context window."
Harness — stateless and decoupled. No longer assumes resources exist in its container. Provisions containers only when Claude requests them via tool calls. Can be rebooted independently without losing session data. Reduced time-to-first-token by ~60% at p50 and >90% at p95.
Sandbox — containers became replaceable "cattle" rather than precious "pets." Each execution environment is a tool with a uniform interface: execute(name, input) → string.
Credential isolation — tokens never become reachable from sandboxes executing Claude-generated code. Two patterns enforce this: resource-bundled auth (tokens used at sandbox init, then discarded) and vault-based OAuth (MCP with tokens stored outside the sandbox).
Scaling — the decoupled design enables multiple stateless harnesses serving the same session, and multiple sandboxes serving the same harness. Customers can run harnesses in their own VPCs without network peering.
Multiagent Orchestration
A Managed Agents session can run more than one agent. The agent the session runs acts as coordinator and is configured with a multiagent block of type multiagent_20261001, which enables three mechanisms: subagents (enabled by default) that take delegated tasks in persistent session threads and can receive follow-ups; dynamic workflows (enabled by default) in which the agent writes a program that runs many agents in phases; and an advisor (off by default) that the primary thread consults mid-turn. Rosters list predefined agents by ID or {"type": "self"}, and inline agents defined at runtime can be disabled.
All agents share one sandbox, filesystem, and vault credentials, but each runs in its own context-isolated session thread; the coordinator reports in the primary thread. Delegation is one level deep, each roster holds up to 20 unique agents, and a session runs at most 25 child threads. The earlier type: "coordinator" configuration is still accepted but documented as legacy. Workflow runs have no approval step, so a session budget should be set at creation. See 📝Claude Coordinator for how this compares with the coordinator in Claude Code Projects.
Messages API vs Managed Agents
Positioned Alongside
- 📝Claude Cowork — Anthropic's enterprise collaboration product
- 📝Claude Code — Terminal-native agent (local execution; Managed Agents is cloud execution)
- 📝Agent SDK — Lower-level SDK for building custom agent loops
Early adopters include Notion, Rakuten, and Asana.
