Claude dynamic workflows Turn One Agent Into a Swarm Boss

Claude dynamic workflows just landed as a beta feature in Claude Managed Agents, and the pitch is pure swarm energy. Instead of one agent grinding through a task, a lead agent writes a small program that fans the work out to many agents in phases and merges what comes back. Anthropic says it crushed a single agent on a bug hunt, but the token bill is the catch.

What actually happened

Anthropic's multiagent docs now describe several ways a managed agent can work with others. Subagents take delegated tasks in their own persistent threads. Dynamic workflows are the new piece: the agent writes a workflow, a program that runs many agents in phases and combines their results, and the server runs it in the background as one workflow run you can follow.

You switch it on by setting the multiagent block's type to multiagent_20261001. By default both subagents and workflows are enabled, and each can use inline agents that the lead defines itself. If you want tighter control over models, prompts and tools, you can list up to 20 predefined agents per list and turn inline agents off.

The docs draw a simple line between the two modes. Subagents fit when the lead needs to follow up with a helper after it reports back, or needs specialists you defined with their own system prompts and tools. Workflows fit most jobs that need more than one agent, such as work with many pieces or work that can run in parallel. You decide what the agent may use and it decides when, so Anthropic suggests spelling out in the system prompt when a workflow run makes sense.

The Decoder covered the launch on October 9 and highlighted Anthropic's benchmark: 70 bugs hidden in a 116,000-line codebase. One agent caught between 14 and 27 per run, while the workflow hit 66. To get started, Anthropic points to the docs or to running "/claude-api managed-agents-onboard" in Claude Code.

The limits that actually matter

Headlines say "1,000 agents in parallel." The workflow runs documentation is more precise. A run can have 64 threads working at once and creates no more until one finishes. Separately, a workflow can start up to 1,000 agents over the run's whole life. Ask for more and the run ends with a thread_limit_error.

So "1,000 in parallel" overstates it. Anthropic also notes the API doesn't guarantee the 64 figure, so it can change. A run lives 24 hours by default, unless the agent sets a shorter lifetime, and a session can hold 10 open runs by default, idle ones included.

Token burn is the real boss fight

Claude dynamic workflows can chew through tokens fast, and The Decoder notes Anthropic itself warns they can use "a lot of tokens." A senior OpenAI engineer recently called agent swarms a massive waste of tokens, which is the bear case in one sentence.

Anthropic's answer is budgets. You set a session budget when you create the session, and it caps spend, runs included. You can't add one to an existing session. When the budget is hit, runs pause, and they resume if you raise or remove it. One more gotcha from the docs: an interrupt stops the agent's turn but doesn't end open runs.

What Polymarket thinks

No market tracks this feature directly. The closest is Which company has the best AI Agent end of October?, which resolves on the top-ranked model in the Agent Arena leaderboard on October 31. At 06:00 UTC on October 10, Anthropic traded near 83.5% (bid 82, ask 85) and OpenAI near 12.5%. That market ranks models, not Managed Agents features, so don't read it as a verdict on workflows.

What to do as a reader

If you build with Claude dynamic workflows, start tiny: a session budget, a small dataset and idempotent tools so retries don't double-write anything. Give the run an explicit time limit in your prompt and check each worker's output instead of trusting a single "completed" status.

If you're just watching the agent race, compare this with our coverage of Claude Opus 5.5 and Claude for Google Workspace. The 66-of-70 result comes from Anthropic's own test, so wait for independent benchmarks before believing swarms beat solo agents everywhere. Nothing here is investment advice.