Claude Managed Agents dynamic workflows: how to use them

Dynamic workflows are a new beta feature in Claude Managed Agents that lets one agent write a small program, called a workflow, which starts many other agents in phases and merges what they return. To use it, you set one field on your agent (multiagent with type multiagent_20261001 and workflows enabled), tell the agent in its system prompt when a job is big enough to warrant a run, and watch the run's progress on the session's event stream.

Anthropic's developer account announced it on October 9. @ClaudeDevs called it a new type of multiagent orchestration, where "a lead agent writes a plan that runs across many agents in phases." The post passed 12,000 likes in a few days, a lot of attention for an API feature.

What @ClaudeDevs posted

The launch thread made two points. The first post said dynamic workflows are now in public beta and are meant for your largest workloads. The follow-up from @ClaudeDevs came with a warning: "Dynamic workflows are powerful and can use a lot of tokens." Anthropic suggests starting with a narrow task and only then making the work harder. If you use Claude Code, the thread points to /claude-api managed-agents-onboard bug-hunter as a starter, and templates are at platform.claude.com/agent-quickstart.

Claude Code got a feature with the same name earlier this year. This launch puts it in Managed Agents, Anthropic's hosted service where the agent runs on Anthropic's servers in place of your laptop.

How to turn it on

Anthropic's API release notes list the setup:

{
  "multiagent": {
    "type": "multiagent_20261001",
    "workflows": { "type": "enabled" }
  }
}

Requests also need the managed-agents-2026-04-01 beta header. Only the agent can start a run, so the system prompt is where you steer it. If you want a workflow for "review every contract in this folder" but not for a single email, write that rule down plainly.

Each run reports its progress through workflow_run.* events: when it's created, when each phase starts and ends, when it pauses, and how it finished.

The limits, from Anthropic's docs

Several write-ups say Claude now runs "1,000 agents in parallel." Anthropic's workflow runs documentation puts it differently:

• Up to 1,000 agents over the whole life of one run. Ask for more and the run ends with a thread_limit_error.
• 64 agents working at once inside a run, a number the docs say isn't guaranteed and can change.
• A 24-hour lifetime per run by default, and 10 open runs per session by default.

So "1,000 at once" overstates it. The real setup is roughly 64 workers at a time, cycling through up to 1,000 tasks.

What it costs

A run has no fee of its own. Every agent's tokens are billed at its model's normal rate, and Managed Agents adds session runtime at $0.08 per session-hour. The token bill is the one to watch. Sixty-four agents each reading a long document add up fast.

There is a brake. Runs count against the session's budget, and when the budget is reached every open run pauses. The docs warn about one gap: each agent finishes the model request it already started, so a run can go past the budget by one request per working agent.

Is it worth it?

Anthropic's own test, reported by The Decoder, planted 70 bugs in a 116,000-line codebase. A single agent found 14 to 27 per run, and the workflow found 66. That's a big jump, but it's a vendor's test on planted bugs, and it doesn't say what each approach cost. A Vals AI comparison we covered found agent teams often cost several times more without scoring much better.

My read: this fits work that splits cleanly into many similar pieces, like checking 400 invoices against purchase orders or reviewing a stack of documents for one clause. It's a poor fit for a task that only needs one careful agent thinking longer.

What this means for a business using AI

For a small business, the part worth noticing is how the work gets split. Big repetitive jobs can now be split up and run overnight without someone writing the orchestration code. The risks are a surprise bill and agents doing more than you meant, so set a budget and keep tool permissions tight before the first run.

If you have a pile of repetitive work and want to know whether it's the kind that splits well, New Face Design offers a free process audit.