Claude Managed Agents
Anthropic service in public beta since April 8, 2026 that runs Claude as an autonomous agent on managed infrastructure. You describe the agent (model, system prompt, tools, MCP servers, skills) + an environment (container with packages, network access), Anthropic handles the runtime: sandbox, persistent sessions, tool execution, prompt caching, compaction, SSE streaming. Built for long-running tasks (minutes to hours) you do not want to operate yourself.
Strengths
- No need to build the agent loop, sandbox, or session management, Anthropic handles it
- Native tools (bash, file ops, web search/fetch) + MCP servers + skills already integrated
- SSE streaming, prompt caching and compaction included, tunable without custom code
- Persistent sessions, file system and conversation history kept across interactions
Limitations
- Beta, behaviors may shift between versions (version header required)
- Less fine-grained control than Messages API over the exact agent ↔ tools loop
- Per-org rate limits (300 create/min, 600 read/min), may constrain multi-tenant scenarios
- Some features (outcomes, multiagent) in research preview, access on request
- Do not believe the carousels announcing '1,000+ startups killed' and 'agents with no backend at all': you still describe the agent, its environment and its tools, and you pay for execution. What goes away is the runtime infrastructure (state, memory, permissions, scheduled runs), not the design work. The early users Anthropic names are well-equipped teams: Notion, Rakuten, Asana, Vibecode, Sentry, Atlassian
Best for
- Product teams that want to expose a Claude agent behind an API without running the infra
- Long async workflows (whole-codebase analysis, data migration, etc.)
- Multi-tenant platforms orchestrating several agents in parallel