Overview
- Anthropic released Claude Managed Agents in public beta on Wednesday, offering a cloud service that runs autonomous AI agents without customers building their own backend systems.
- A production agent typically needs sandboxed code execution, checkpointing to save state, credential and permission controls, and end-to-end tracing, and the new runtime bundles these pieces so teams do not assemble them from scratch.
- Developers define an agent’s tasks, tools, and guardrails and then run it on Anthropic’s infrastructure, which supports hours-long sessions, secure tool use, error recovery, and dashboards to monitor what the agent is doing.
- Pricing adds a runtime fee of $0.08 per session-hour on top of standard model token costs, with web search billed at $10 per 1,000 searches, and idle wait time does not count toward the runtime charge.
- Several capabilities remain in limited research preview, including advanced memory tooling, multi-agent coordination, and self-evaluation loops that let agents iterate until they meet a defined goal.