Claude Managed Agents
Also known as: Anthropic Managed Agents, Managed Agents by Anthropic, Claude Managed Agents API, Self hosted Claude Managed Agents
Claude Managed Agents is Anthropic's hosted harness for running Claude as an autonomous agent through the Claude API. Developers define an agent's model, prompt, tools, MCP servers and skills once, then start long running sessions in Anthropic's cloud sandboxes or their own, with memory stores across sessions, per user credential vaults, approval policies, rubric graded outcomes, subagents and cron schedules. It bills $0.08 per running session hour plus Claude tokens and is in public beta.
Claude Managed Agents is Anthropic's hosted service for running Claude as an autonomous agent, launched in public beta on April 8, 2026 as part of the Claude Platform. Where the Messages API gives developers raw model calls, Managed Agents supplies the agent loop, the sandboxes and the stored state, so a team defines an agent once and starts sessions against it over the API. Every Claude API account has access by default, and the service also runs on Claude Platform on AWS and on Claude models on Google Cloud.
An agent bundles a model, a system prompt, tools, MCP servers and skills, and an environment sets where its sessions run: an Anthropic managed cloud sandbox with packages and network access, or a sandbox on the customer's own infrastructure. Inside a session Claude runs bash, edits files, searches and fetches the web within allowed domains, calls MCP servers and hands custom tools back to the application, streaming each step as an event while the service keeps the full history. Sessions run for hours and accept new instructions or an interrupt midway.
Outcomes attach a rubric so a separate grader scores the work and sends it back for revision. In multiagent mode one agent delegates to as many as 20 subagents, runs dynamic workflows of many agents in phases or consults an advisor model. Memory stores carry lessons from one session to the next with every change versioned, vaults hold each end user's OAuth tokens and secrets, permission policies decide which tool calls wait for approval, and scheduled deployments start runs on a cron schedule.
Pricing is usage based. A session costs $0.08 per hour while it is running, plus Claude tokens at standard rates, from Haiku 5.5 at $0.10 per million input tokens to Opus 5.5 at $4 and Fable 5.2 at $8, with web searches at $10 per 1,000. Pinning inference to the United States multiplies token prices by 1.1.
Teams work through Claude Console workspaces and workspace scoped API keys under six organization roles, and Anthropic keeps a SOC 2 Type 2 report and an ISO 27001 certificate in its trust center.
The hosted service stores transcripts until they are deleted, which places it outside zero data retention and HIPAA coverage, while a self hosted edition sold through Anthropic's account team runs the whole service in the customer's cloud account or data center.
It fits product and platform teams that want long running agents inside their own products without building a loop, sandbox and state layer. Rakuten put specialist agents for product, sales, marketing and finance into Slack and Teams, each deployed within a week, Asana built its AI Teammates on the service, Sentry shipped an agent that writes patches and opens pull requests from Seer's findings in weeks instead of months, and Vibecode spins up app infrastructure at least ten times faster. It suits teams less well when they need models from other providers, inbound event triggers without their own glue code, or HIPAA and zero retention coverage on the hosted service.
Vendor details
Canonical URL
https://platform.claude.com/docs/en/managed-agents/overview
Category
Agent infrastructure
Subcategory
Anthropic's hosted agent harness and sandboxes for long running Claude agents, run through the API with sessions, memory stores, vaults and outcomes
Funding status
A product of Anthropic.
Company status
first party product
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Built in bash, file, web search and web fetch tools; remote MCP servers authenticated through vaults with OAuth refresh; custom tools executed by the customer's application; GitHub access and file mounts through the Files API; Agent Skills; outbound signed webhooks; Claude models through the Claude API, Claude Platform on AWS and Claude models on Google Cloud; self hosted sandboxes and a self hosted service on Docker Compose or Kubernetes.
In practice
A project management company gives each customer workspace an agent that picks up assigned tasks. Each end user's OAuth tokens sit in a vault, sessions run for an hour in Anthropic's sandbox, and a memory store per team keeps conventions the agent learned last week.
A finance team sends one event asking for a DCF model of a retailer in a spreadsheet, with a rubric file attached. The agent builds the model, a separate grader marks two criteria as failing, and the agent revises until the grader returns satisfied.
A platform team sets a scheduled deployment that runs every weekday at 6am New York time, starting a session that triages overnight error reports, delegates fixes to subagents and posts a summary, with a webhook telling the team's service when the run goes idle.
Sources & related URLs
Related / legacy domains
Research sources
Agentic Index coverage score
11.0 / 14 capabilities · 79%
| Integrations & Tool Calling | Full |
|---|---|
|
Each agent starts with a toolset of bash, read, write, edit, glob, grep, web search and web fetch, all running in its sandbox. Remote MCP servers connect outside systems, with credentials drawn from a per user vault that refreshes OAuth tokens on its own, and custom tools let the agent ask the customer's application to run an operation and return the result. GitHub repositories and uploaded files mount into the session, and Agent Skills add packaged instructions and scripts for particular tasks. Sourceplatform.claude.com/docs/en/managed-agents/toolsread 2026-11-19 |
|
| Workflow Orchestration | Full |
|
In multiagent mode a lead agent hands work to subagents in context isolated threads that share the sandbox and vault credentials. It can delegate to up to 20 predefined agents or define its own, send follow ups to a subagent's persistent thread, write a program that runs many agents in phases as a background workflow, or consult an advisor model in the middle of a turn. A session holds at most 25 child threads, delegation goes one level deep, and a session budget caps the tokens a workflow can spend. Sourceplatform.claude.com/docs/en/managed-agents/multi-agentread 2026-11-19 |
|
| Knowledge Grounding & RAG | Partial |
|
Agents ground their work in what a session brings with it: files uploaded through the Files API and mounted into the sandbox, cloned GitHub repositories, skills, MCP servers that reach the customer's own data, and web search and fetch limited to allowed domains, with grep and glob for searching inside the sandbox. Context is gathered for each run. There is no managed index, vector store or retrieval layer that keeps the customer's documents searchable between sessions. Sourceplatform.claude.com/docs/en/managed-agents/overviewread 2026-11-19 |
|
| Human Oversight & Guardrails | Full |
|
Permission policies govern every built in and MCP tool: a tool can run automatically, wait for the application to approve each call, or go to the server, which evaluates the call before it runs. While a session works, the application can send new user events to steer it or an interrupt to stop and change direction, and memory stores can be attached read only so a prompt injection cannot rewrite them. Custom tools stay under the customer's own control because the application executes them. Sourceplatform.claude.com/docs/en/managed-agents/permission-policiesread 2026-11-19 |
|
| Security, Identity & Governance | Full |
|
Access runs through the Claude Console's role based system of six organization roles, from User to Admin, with workspaces that scope API keys, members and spend limits and an Admin API for managing them. Vaults keep each end user's OAuth tokens, bearer tokens and secrets apart, send them only to allowed hosts and inject environment secrets at egress so the agent never sees them. A SOC 2 Type 2 report and an ISO 27001 certificate sit in the Anthropic Trust Center. The hosted service is not eligible for zero data retention or HIPAA coverage. Sourcesupport.claude.com/en/articles/10186004-api-console-roles-and-permissionsread 2026-11-19 |
|
| Observability & Auditability | Full |
|
Every session streams its work as server sent events: agent messages, tool calls and results, status changes and span events, including each outcome evaluation with the grader's explanation and token usage. The service stores the full event history server side for retrieval at any time, so a team can replay what an agent did and why. Signed webhooks report run starts, idles, terminations, budget limits and changes to vault credentials, and each memory edit leaves an immutable version behind for audit or rollback. Sourceplatform.claude.com/docs/en/managed-agents/overviewread 2026-11-19 |
|
| Memory & State Persistence | Full |
|
Memory stores, in beta, carry preferences, project conventions and past mistakes from one session to the next. They are workspace scoped, up to eight attach to a session so a team can map stores to end users, teams or projects, and each mounts as a folder the agent reads and writes with its file tools, or read only. Stores last until deleted or archived, every change is kept as a version for 30 days, and memories can be viewed, edited or deleted in the API or Console. Dreams, which tidy a store from past transcripts, sit behind a research preview. Sourceplatform.claude.com/docs/en/managed-agents/memoryread 2026-11-19 |
|
| Deployment & Data Residency | Full |
|
Sessions run in an Anthropic managed cloud sandbox or in self hosted sandboxes on infrastructure the customer controls. Through Anthropic's account team, the whole service, including the API, agent loop and storage, can run in the customer's own cloud account or on premises data center from a Docker Compose stack or Kubernetes, with models served by Anthropic, Amazon Bedrock, Google Cloud, Microsoft Foundry or the customer's gateway. The hosted service also runs on Claude Platform on AWS and Claude models on Google Cloud, and inference can be pinned to the United States. Sourceplatform.claude.com/docs/en/managed-agents/self-hostedread 2026-11-19 |
|
| Prebuilt Agents / Templates / Packs | Partial |
|
Anthropic ships a ready harness rather than ready agents. The prebuilt toolset, prompt caching, compaction and the cloud sandbox with preinstalled packages come configured, and Agent Skills package instructions and scripts the agent can load, but every agent's prompt, tools and job are written by the customer. There is no gallery of finished agents to switch on, and the cookbook recipes and quickstart serve as examples to adapt rather than products. Sourceplatform.claude.com/docs/en/managed-agents/overviewread 2026-11-19 |
|
| Triggers & Channel Coverage | Partial |
|
Work arrives two ways. The customer's application creates a session and sends events over the API, and scheduled deployments start sessions on their own from a POSIX cron expression in a chosen time zone, up to 1,000 schedules per organization, with a manual run endpoint beside them. Webhooks travel outward only, reporting status changes, budget limits and finished runs, so an inbound email, chat message or record change reaches an agent only through code the customer writes to call the API. Sourceplatform.claude.com/docs/en/managed-agents/scheduled-deploymentsread 2026-11-19 |
|
| Model Flexibility & Routing | Not documented |
|
Only Claude runs here. A customer picks which Claude model each agent uses, from Haiku 5.5 to Opus 5.5, Fable 5.2 and Mythos 5.2, can switch on fast mode, and in multiagent mode can add an advisor model the lead thread consults, but every option comes from Anthropic. No model from another maker can be selected, and the routes through Amazon Bedrock, Google Cloud or Microsoft Foundry in the self hosted edition still serve Claude. Sourceplatform.claude.com/docs/en/managed-agents/overviewread 2026-11-19 |
|
| APIs / SDKs / MCP Extensibility | Full |
|
Everything is an API. REST endpoints create agents, environments, sessions, deployments, vaults and memory stores, send and stream session events, and list output files, all under one beta header. Official SDKs in Python, TypeScript, C#, Go, Java, PHP and Ruby set that header automatically, and the ant command line tool covers the same operations from a shell. Agents are created once and referenced by ID across sessions, so an application can manage its whole fleet in code. Sourceplatform.claude.com/docs/en/managed-agents/define-outcomesread 2026-11-19 |
|
| Testing, Debugging & Optimization | Full |
|
Outcomes turn a session into a graded loop. The customer writes a rubric of explicit criteria, and the harness provisions a grader in a separate context window that scores the agent's artifact against each one and returns an explanation of what passed and failed. A needs_revision result sends the agent into another iteration, up to 20, and each evaluation is logged with its result, explanation and token usage, so a team can see how often an agent meets its bar. Sourceplatform.claude.com/docs/en/managed-agents/define-outcomesread 2026-11-19 |
|
| Browser / Computer-use | Partial |
|
Web access comes through web search and web fetch, which can be held to an allowlist or blocklist of domains, capped in content size and set to a search location. Those tools read pages rather than drive them. The toolset has no browser the agent clicks through and no desktop or computer use tool, so form filling on a live site needs an MCP server the customer adds, while bash and code run inside a sandbox that has no screen. Sourceplatform.claude.com/docs/en/managed-agents/toolsread 2026-11-19 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
$0.08 per running session hour plus Claude tokens, from Haiku 5.5 at $0.10 input and $0.50 output per million; web search $10 per 1,000
Session runtime per hour while running, plus model tokens and per use tool fees
What is public
Claude Managed Agents charges $0.08 per session hour while a session is running, plus Claude tokens at standard rates. Haiku 5.5 costs $0.10 input and $0.50 output per million tokens on prompts up to 100,000 tokens, Sonnet 5.5 $2 and $10, Opus 5.5 $4 and $20, and Fable 5.2 $8 and $40.
Billing mechanics
Runtime accrues only while a session is running and stops while it is idle, rescheduling or terminated, and it replaces separate container hour billing for code execution. Prompt caching discounts apply to tokens, while the Batch API discount does not, and web searches cost $10 per 1,000.
Cost watchouts
Multiagent workflows run many agents at once, each using tokens, and every outcome check adds a second model's tokens to each round of revision. Pinning inference to the United States multiplies token prices by 1.1.
Variable cost rationale
Long sessions bill runtime by the hour and tokens on every turn, and multiagent workflows multiply both, so spend tracks how much work the agents do.
Additional watchouts
Fast mode carries premium token rates, and setting inference to the United States multiplies token prices by 1.1.
Sales call required
Mixed (some tiers require a call)
Free / trial
Included with every Claude API account, which starts with free test credits
Lowest paid plan
No plan; pay as you go at $0.08 per running session hour plus Claude tokens
Commercial notes
Access comes with every Claude API account during the public beta. The self hosted edition, which runs the whole service on the customer's own infrastructure, is arranged through Anthropic's account team.
Missing data
Anthropic gives no price for the self hosted edition.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…