Claude Managed Agents Draft
Anthropic · reviewed 22 November 2026 · its card
Claude Managed Agents lets a developer run Claude as an autonomous agent for long-running and asynchronous work without building an agent loop, tool execution or runtime of their own: it is a "pre-built, configurable agent harness that runs in managed infrastructure" (S1). Anthropic makes it, and sets it beside the Agent SDK as "a hosted agent harness that runs the agent loop, with sessions in an Anthropic-managed cloud sandbox or a self-hosted sandbox on your own infrastructure", used "from the SDK for your language, the ant CLI, or the REST API" (S2). It runs on the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed), all in beta behind the managed-agents-2026-04-01 header ("Managed Agents are in beta on every model", so its rows are partial), and as self-hosted Claude Managed Agents on the customer's own infrastructure, by request (S1, S13). It is a paid service, billed on tokens at model rates and on session runtime at $0.08 a session-hour; no licence is stated on the pages read (S20). It launched in beta on 8 April 2026, and the latest change seen is the self-hosted service of 12 November 2026; the API carries no product version number (S4). It is an agent harness and the agent host of its sessions: "The Managed Agents service keeps the event history" (S1). Its maker's own words for it are on the vendor card Agentic harness (Anthropic).
- An instance of
- Agent harness, Agent host
Capabilities
| Capability | Value | Detail | Sources |
|---|---|---|---|
| Model | Partial | Beta. Claude only: "Claude 4.5 and later models are supported"; a self-hosted server can send model calls to Amazon Bedrock, Google Cloud, Microsoft Foundry or "a model gateway that you run" | S5 S14 |
| Session | Partial | Beta. Sessions "resume cleanly after pauses, and store conversation history, sandbox state, and outputs server-side"; a session "maintains conversation history across multiple interactions" and is archived or deleted through the API | S1 S6 S7 |
| Subagent | Partial | Beta. "With subagents, it delegates tasks itself and reads what each subagent reports", each in its own session thread, one level deep; dynamic workflows run many agents in phases | S8 S9 |
| Agent team | Unknown | "A lead agent assigns work to specialists" sharing a filesystem; the docs show subagents reporting only to the primary thread, not to each other | S3 S9 |
| Skill | Partial | Beta. "A custom skill is a directory containing a SKILL.md file", uploaded or found in a mounted repository's .claude/skills; the pages read link Anthropic's Agent Skills docs but do not name the specification |
S10 |
| Instruction file | Unknown | Not found; the agent's system prompt is set on its definition | S5 |
| Plugin | Unknown | Not found | None |
| MCP | Partial | Beta. Remote MCP servers declared "by name and URL", authenticated from a vault; private servers through MCP tunnels, a research preview | S11 S1 |
| Execution environment | Partial | Beta. vendor-sandbox: "Each session gets its own fresh Linux container on Anthropic-managed infrastructure"; or a self-hosted sandbox, "containers, VMs, or other isolated workspaces" the customer runs |
S12 S13 |
| Client | Partial | Beta. The Console's session viewer and inline session runner; ant beta:sessions connect attaches a terminal to a session; SDKs and REST API. No desktop or mobile window found |
S18 S19 S4 |
| Agent host | Unknown | Hosts sessions of Anthropic's own harness only; nothing found on other makers' harnesses | S1 |
| Channel | Unknown | The client's own events can "guide the agent while it runs"; no outside-event channel documented, and webhooks only notify outward | S1 S17 |
| Cross-session messaging | Unknown | Messages pass between threads of one session; memory stores are shared across sessions, but no messaging between sessions found | S9 S15 |
| ACP | Unknown | Not found | None |
| AHP | Unknown | Not found | None |
| A2A | Unknown | Not found | None |
| Bot | Partial | Beta, in part: a named, versioned agent with memory stores that "carry information across sessions", and scheduled deployments that "start sessions autonomously" on a cron; no mail to the agent or event wake documented | S5 S15 S16 |
State of play
How things stand, newest first, as the card records it.
- Self-hosted Claude Managed Agents launched: the whole service, with agents, sessions and session history, on the customer's own infrastructure, by request to the account team. S4 S13
- Dynamic workflows, in beta: the agent writes a program that "runs many agents in phases". S4
- Permission policies gain
auto;antCLI 1.32.0 addsant beta:sessions connect, attaching a terminal to a session. S4 - Scheduled deployments run sessions on a cron schedule. S4
- Webhooks, multiagent orchestration and self-hosted sandboxes reach Claude Platform on AWS. S4
- Dreams, which reorganise a memory store from past transcripts, arrive as a research preview. S4
- Memory in beta under the standard header. S4
- Claude Managed Agents launched in beta, "a fully managed agent harness for running Claude as an autonomous agent". S4