ooligo

Claude Managed Agents

ai-agent-platform ai-agents · agent-orchestration · llm-platform · sandboxed-execution
AI-NATIVE MCP API
RevOpsLegal OpsRecruiting & TA
8.0 /10

What it is

Claude Managed Agents is Anthropic’s hosted runtime for long-running agents, part of the Claude Platform (the API). You define an agent (model, system prompt, tools, MCP servers and skills), an environment (the sandbox it works in), and start a session, which is a running instance of that agent. Anthropic runs the agent loop, the container and the tool calls; you send user turns and receive events over a stream. It launched in public beta on 8 April 2026, and every endpoint still requires the managed-agents-2026-04-01 beta header, which the SDKs set for you. Access is on by default for all API accounts.

Each session gets its own fresh Linux container with common languages and package managers preinstalled. The built-in toolset covers bash, file read, write and edit, glob, grep, web_fetch and web_search, all enabled by default, and each tool can carry a permission policy that pauses for approval. Remote MCP servers connect over streamable HTTP. Credentials live in vaults: OAuth tokens for MCP servers with automatic refresh, bearer tokens, and environment variables that the sandbox only sees as placeholders, swapped for the real secret at egress. Since May 2026, outcomes (a rubric the agent grades itself against), multi-agent orchestration, memory and webhooks are in public beta; dreaming and MCP tunnels to private servers are a narrower research preview. On 12 November 2026 Anthropic opened a private beta of self-hosted Managed Agents, which runs the whole service in your own cloud.

The question it answers for an ops team: do we run headless agents (a nightly CRM hygiene sweep, a contract-intake triage, a candidate-digest agent) on Anthropic’s runtime, or keep self-hosting the Agent SDK on n8n or our own workers?

What it costs in practice

Billing has two lines. Tokens are charged at the standard Claude API rates for the model you pick, with prompt-caching multipliers applied as usual. Runtime is $0.08 per session-hour, measured to the millisecond and charged only while the session status is running. Time spent idle, waiting for your next message or a tool approval, is not billed, and session runtime replaces the separate code-execution container-hour charge. Web search inside a session costs $10 per 1,000 searches. There is no seat fee or platform fee, and usage counts toward an existing Anthropic commit.

Anthropic’s own worked example, a one-hour Opus 5.5 session with 50,000 input and 15,000 output tokens, comes to $0.58: $0.20 input, $0.30 output, $0.08 runtime. With 40,000 of those input tokens served from cache it drops to $0.428. The lesson for budgeting is that runtime is the small line. A session that genuinely ran around the clock for a month would add about $58 in runtime; a nightly sweep that runs 20 minutes adds under $1 a month. Model tokens are what to forecast and cap. Two discounts do not apply: there is no Batch API pricing for sessions, and partner-cloud pricing does not exist because the hosted runtime is not sold through Bedrock or Vertex.

Best for

  • RevOps or Legal Ops teams with one engineer who want an agent that runs for an hour in a sandbox, calls the CRM or CLM through MCP, and reports back, without standing up a container fleet, a job queue and a secrets store first.
  • Teams already on a Claude API commit building an agent into an internal tool or product, where the runtime fee and the console tracing replace weeks of orchestration code. Anthropic’s launch customers (Notion, Rakuten, Asana, Sentry) report shipping in “weeks instead of months”; these are vendor-published figures.

Alternatives, and when to pick them instead

  • Amazon Bedrock with AgentCore — one of the two market-share leaders among cloud agent platforms. Pick it when identity, data and the cloud commit live in AWS and the agent must run in your own account. Note that Claude on Bedrock lacks the Managed Agents runtime entirely.
  • Microsoft Foundry Agent Service — the other leader. Pick it when you need Azure tenancy, Entra identities for agents and a choice of models beyond Claude.
  • Self-hosting the Claude Agent SDK on n8n or your own workers — pick it when you need zero data retention, HIPAA eligibility or a hard session-length guarantee today. You own the sandbox, retries and secrets, and pay only tokens.
  • Claude Cowork — pick it when the user is a person at a desk, not a service. Cowork is the seat-priced desktop agent; Managed Agents is the API runtime for headless work.
  • Hermes Agent — the fastest-growing entrant on the self-hosted side, at roughly 263,000 GitHub stars nine months after release. Pick it for one resident agent in Slack or Telegram on any model, with no API integration work.

Watch-outs

  • Network egress is open unless you close it. An environment created through the API without a networking setting defaults to unrestricted (the Console form starts on Limited). Guard: set networking to limited with an explicit allowed_hosts list in every environment you create from code, and remember that list also governs web_search and web_fetch.
  • Not eligible for ZDR or HIPAA. Anthropic’s docs say the hosted runtime is outside zero data retention and HIPAA, across all sub-features. Guard: keep PHI and anything under a ZDR clause on a self-hosted Agent SDK build, or wait for the self-hosted private beta, which is the one variant ZDR covers.
  • It is a beta with a dated header. Endpoints, event types and defaults can change, and the docs publish no maximum session length or concurrency limit; the published rate limits are 300 create and 1,200 read requests per minute per organization. Guard: pin SDK versions, write the agent and environment definitions to files you apply with the ant CLI so they live in git, and set organization spend limits before the first scheduled run.
  • Memory persists between sessions. Memory stores carry over from run to run, so a wrong fact the agent saves keeps steering later runs. Guard: give each production agent its own memory store, read its contents through the API on a fixed schedule, and delete entries that no reviewed source supports.