Claude Dynamic Workflows come to Managed Agents

Claude Managed Agents gets dynamic workflows in public beta

Claude Managed Agents is Anthropic's service for running AI agents on its own servers. An agent here means a model that works through a task by itself, reading files and running commands along the way.

The October 9 release notes say these agents can now use dynamic workflows. The feature is in beta, so API calls need the managed-agents-2026-04-01 beta header.

A dynamic workflow is a program that runs many agents

A workflow is a program that splits a job into phases, runs many agents in each phase, and combines their results. You do not write it.

The agent you gave the task to writes it, and the server then runs it in the background as a single "workflow run." The main agent can keep working or end its turn while the run is going.

The example in Anthropic's docs is a contract review. You ask which of 300 contracts contain a change-of-control clause. One phase splits the contracts across many agents that read them in parallel, a second phase merges their findings, and the agent comes back with one answer: 41 of the 300 contracts have the clause.

Audits, migrations and deep research are the target jobs

Workflows pay off when the same steps have to be applied to a large number of items. Anthropic lists audits, migrations, deep research and cross-checking, where a second set of agents verifies the first set's work. Reviewing hundreds of documents is the same kind of job.

For a job with only one or two items, a workflow is overkill. Anthropic's own contract example tells the agent to read one or two contracts itself and to save the workflow run for larger batches.

Claude Code has had dynamic workflows since May

Dynamic workflows started as a Claude Code feature, announced on May 28. Anthropic has since updated that blog post to say the feature is generally available. In Claude Code it is available on all paid plans, and it also works through the Anthropic API, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry.

The headline example in that post is a port of Bun, the JavaScript runtime, from Zig to Rust. According to the post, the result was about 750,000 lines of Rust, eleven days from first commit to merge, and 99.8% of the existing tests passing. Claude Opus 4.8: Pricing, Fast Mode, Workflows covers the Claude Code launch that came with it.

Subagents run a few tasks, dynamic workflows run hundreds

The practical difference is scale: subagents handle a few tasks per turn, while a workflow runs dozens to hundreds of agents. The gap comes from where interim results are kept, and that depends on who plans the work.

With dynamic workflows, the script holds the plan

With subagents and agent teams, Claude itself decides, turn by turn, what to hand out next. Every result lands in Claude's context window, the amount of conversation the model can hold at once. The more items you add, the faster that window fills up, which is why subagents top out at a handful per turn.

A workflow moves the plan into code. The script owns the loop, the branching and every intermediate result, and only the final answer reaches Claude's context. That is what lets one run go to dozens or hundreds of agents.

Claude dynamic workflows resume after an interruption

If a subagent turn is interrupted, the turn starts again from the top, but a workflow picks up where it stopped within the same session. The script is also saved as a file in the session folder, so you can read it, edit it, or keep it for the next similar job.

Running many agents side by side brings its own failure modes, such as agents drifting toward the same wrong conclusion. Anthropic's research on that is covered in Multi agent AI sabotage in Anthropic's test.

Managed Agents has subagents too. Delegation goes one level deep, and a session can hold up to 25 subagent threads at a time, idle ones included. Threads inside a workflow run do not count toward that 25.

Claude Code workflows vs Managed Agents workflows

The feature is the same in both products, but where it runs and who starts it are different. In Claude Code, the workflow runs on your own machine and starts when you ask for it. In Managed Agents, it runs on Anthropic's servers and the agent decides when to start a run.

Claude Code runs a workflow on your machine when you ask

In Claude Code, a workflow is a JavaScript program that Claude writes and a runtime executes in the background on your machine. Your session stays free while it runs, so you can keep talking to Claude. The workflow itself never stops to ask you for input; it only waits on permission prompts and usage limits.

A Managed Agents agent decides when to start a run

Managed Agents has no separate API call for starting a workflow run. The agent reads each incoming message and decides whether to start a run, and when. You control that choice through the system prompt, where you tell the agent which kinds of requests deserve a run.

All agents in a session share one sandbox, one file system and one set of credentials, while each thread keeps its own conversation history. Once a thread inside a run has started, you cannot send it further instructions.

Managed Agents enables workflows in the multiagent block

You turn the feature on in the multiagent block of the agent definition. Set type to multiagent_20261001 and set workflows to enabled.

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

The multiagent_20261001 type also turns on subagents

Under this type, both workflows and subagents are on by default. The advisor setting, which lets the agent ask another model for advice, is off by default. If you only want workflows, set subagents to disabled.

Put the trigger for a workflow run in the system prompt

The system prompt is where you tell the agent when a workflow run is worth starting. Anthropic's example is a contract-review agent that changes its behavior based on how many contracts it receives.

Writing a threshold like this for your own workload stops a small request from fanning out to dozens of agents. It is the first and cheapest way to keep token use in check.

Register predefined agents to use another Claude model

Agents in a run are either defined inline by the workflow or taken from a list you register. Inline agents use the same model as the session. To run part of the job on a different model, add that agent's definition to workflows.predefined_agents, which holds up to 20 entries.

Progress reaches the session's event stream as events whose names start with workflow_run. You can also list and open the thread of each agent inside the run.

In Claude Code, start with /deep-research or ultracode

Claude Code gives you three ways to start a workflow. On the Pro plan, first switch the feature on from the Dynamic workflows row in /config.

1. Try /deep-research first. It is a workflow built into Claude Code. Give it a question and it searches the web from several angles, cross-checks the sources it finds, and returns a cited report with unsupported claims filtered out.
2. Add ultracode to a single prompt. This runs one task as a workflow without changing the session's effort level. Asking in plain words, such as "use a workflow," counts the same way. Name the folder and the checks you want so the run stays small.
3. Set /effort ultracode for the whole session. Claude then decides task by task whether to plan a workflow, and one request can split into separate workflows for understanding, changing and verifying the code. Starting Claude Code with claude --effort ultracode sets the effort level to xhigh.

Anthropic's blog post recommends turning on auto mode before you use workflows, so that fewer permission prompts interrupt a run. What auto mode approves on its own is covered in Claude Code Auto Mode Is Now the Default Setting.

Claude dynamic workflow limits and costs

The 1,000-agent limit is the total a run can start over its lifetime, not the number running at once. To keep token spending under control, Managed Agents gives you a session budget, and Claude Code relies on you keeping the job small.

1,000 agents per run, 64 running at once

In Managed Agents, a run works on up to 64 threads at a time and can start 1,000 agents in total. The other limits are a lifetime of 24 hours by default and 10 open runs per session by default, paused runs included.

Claude Code runs 16 agents at once by default, fewer on machines with few CPU cores. From v2.1.269 you can set CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS anywhere from 1 to 256.

The 1,000 total is there to stop runaway loops. In Managed Agents, the server may retry a failed agent on a new thread, so the thread count can end up above 1,000.

Token costs and session budgets in Managed Agents

A workflow run has no fee of its own; you pay for the tokens its agents use, at each model's rates. Those tokens are billed the same way as the rest of the session. A session budget caps that spending, and you can only attach one when you create the session.

When the budget is reached, every open run pauses, and the runs resume once you raise or remove the budget. Model calls inside a run also count against your organization's Messages API rate limits. Because a run can create more than one thread for the same piece of work, Anthropic asks you to make the tools your agents call safe to call twice.

Claude Code warns before a large workflow

In Claude Code, a workflow can use noticeably more tokens than doing the same job in conversation, and it all counts toward your plan's usage and rate limits. When a workflow plans more than 25 agents, or its estimated total passes 1.5 million tokens, a Large workflow warning appears below the prompt.

The warning does not block the run. Anthropic suggests starting with a narrow task, and pointing a run at one folder instead of the whole repository is a good way to learn how much a workflow costs you.

Save dynamic workflows for jobs with many items

Claude Dynamic Workflows let a program that Claude writes run many agents in parallel and merge their results. Because interim results stay in the script, not in Claude's context, a single run can reach dozens or hundreds of agents, far more than subagents. Since October 9 the feature is also in public beta in Claude Managed Agents, where runs happen on Anthropic's servers and the agent decides when to start one.

In Claude Code, /deep-research or a tightly scoped ultracode prompt is the easiest way to try it. In Managed Agents, create the session with a budget and write the conditions for starting a run into the system prompt.