Claude Managed Agents vs Cowork Plugins
“Managed agents” is easy to confuse with Cowork’s role plugins. On Anthropic’s Platform, Claude Managed Agents is a beta API harness for long-running agents in managed sandboxes. In Claude Desktop / Cowork, “agents” for finance or legal work usually means plugins and skills—not that same API product.
This page only states what we can verify from Anthropic documentation. We do not repeat third-party marketing claims (launch slogans, accuracy percentages, or undated competitor copy) unless the same claim appears in an official Anthropic page we checked.
Two different products (keep them separate)
| Claude Managed Agents (Platform API) | Cowork plugins / vertical agents (Desktop & Cowork) | |
|---|---|---|
| What it is | Pre-built, configurable agent harness on Anthropic-managed (or self-hosted) infrastructure | File-based plugins that bundle skills, connectors, slash commands, and related workflow pieces |
| Who it is for | Developers building apps with the Claude API | Knowledge workers and teams using Claude Cowork / Desktop |
| Where it runs | Cloud sandbox or self-hosted environment you configure via API | Cowork / chat surfaces after you install plugins from a marketplace |
| Official entry | Managed Agents overview | Use plugins in Claude; finance set documented at FSI plugins |
Some Anthropic legal materials mention that certain practice-area cookbooks can also be deployed as Managed Agents on the Platform. That does not make every Cowork plugin “Managed Agents API.” When in doubt, check whether you are holding an API key and creating agents / sessions, or installing a marketplace plugin in Desktop.
What is verified for Managed Agents (API)
From the official overview (beta):
- Provides harness + infrastructure so you do not build your own agent loop, tool execution, and runtime.
- Core concepts: Agent, Environment, Session, Events.
- Environments may be an Anthropic-managed cloud sandbox or a self-hosted sandbox.
- Built-in tool access described includes bash, file operations, web search/fetch, and MCP servers.
- Best suited to long-running / asynchronous work, stateful sessions, and scheduled deployments (see Platform docs for those subpages).
- Requires the
managed-agents-2026-04-01beta header on Managed Agents endpoints (SDKs set it automatically). Behaviors may change while in beta. - Stateful by design; the overview states Managed Agents is not currently eligible for Zero Data Retention or HIPAA BAA coverage—confirm on Anthropic’s API and data retention docs before regulated workloads.
We intentionally omit competitor-style metrics that do not appear on the overview page we verified.
What “agents” means inside Cowork
In Cowork, specialized “agents” for a function are usually plugins:
- Role packs (productivity, legal, finance, etc.) from Anthropic or partners
- Vertical finance plugins documented in finance plugins
- Legal role plugin documented in legal plugins
- Admin-controlled marketplaces described in Claude Help
If your goal is “run IB / equity research / in-house legal workflows in Desktop,” start with plugins—not the Managed Agents API.
When to choose which
- Choose Cowork plugins when a human drives work in Claude Desktop / Cowork and needs packaged skills + connectors for a role.
- Choose Managed Agents API when you are building a product or backend that must run long autonomous sessions in a sandbox with API control.
- Choose Claude Code Review when you need the managed GitHub PR review service—that is a separate Claude Code preview (Code Review), not Managed Agents and not a Cowork plugin.
Honest gaps
Public docs for Managed Agents are solid on concepts and beta access, but product marketing pages on third-party sites often blend API agents, Cowork plugins, and Code Review. If a claim is not on Anthropic’s Platform docs or Claude Help / blog posts linked here, treat it as unverified.
Related Guides
- Finance plugins — Official FSI plugin set and marketplace install
- Legal plugins — Official Legal plugin, slash commands, and not-legal-advice limits
- Cowork plugins — How plugins work in Desktop
- Connectors hub — Connectors vs MCP vs plugins
- Security model — Cowork sandbox / isolation framing
- Sandbox feature pointer — Short link hub for sandbox search intent
Related Features
- Code Review — Managed multi-agent PR reviews (Claude Code)
- Scheduled Tasks — Recurring Cowork runs
- Dispatch — Mobile → Desktop handoff