Anthropic’s claude-api skill now builds Managed Agents that run dynamic workflows

Anthropic's Claude API skill can now build Managed Agents that run dynamic workflows. It merged on October 9th. Ask it for an agent whose job is too big to hand out one task at a time, and it switches workflows on.

With workflows on, the agent writes a plan that runs many agents in phases. The server runs it in the background, so the agent can keep working. It's separate from Claude Code's own Workflow tool.

The skill decides by the size of the job. A roster of subagents tops out at 25 child threads. A job bigger than that gets a workflow run, which today keeps up to 64 agents working at once.

The skill sets the new multiagent type with workflows enabled. The agent decides when to start a run, so the skill also writes a system prompt line saying when. That line stays conditional, because every agent in a run uses tokens.

When the skill turns workflows on, its reply says so. It says every agent uses tokens, so the session needs a budget, and shows how to switch workflows off. If you've said no, it builds a plain roster instead.

The skill gains four quickstart templates built on workflows, for 13 in all. A bug hunter reports only bugs a separate agent proved, and a performance tuner optimizes every slow function. There's also a video maker and a stock watchlist scanner.

After building a large Messages API job, the skill now offers a Managed Agents version, in a few sentences with no dollar figures. It offers once per conversation, and skips small jobs or anyone who asked for the Messages API by name.

The skill's networking advice is corrected. Under limited networking, web search and web fetch reach only the allowed hosts, and with none listed they return nothing. Unrestricted networking opens the whole sandbox, so the skill offers it with that warning and lets you choose.

Managed Agents is still in beta. A workflow run lasts 24 hours by default, a session holds 10 open runs, and a run past about 1,000 threads ends with an error. The change is pull request 1995 in Anthropic's skills repository.