Canonical: https://botbento.com/blog/claude-managed-agents-dynamic-workflows/ Format: Markdown representation of the public HTML page. [Home](/) / [Blog](/blog/) FIELD NOTES / 4 MIN READ # Claude Managed Agents Can Write Its Own Workflow. Should Yours Use One? Anthropic's Claude Managed Agents can now write and run its own multi-agent workflow instead of delegating to subagents one call at a time. This piece walks through the exact config that turns it on, the documented limits that bound a run, and the checklist to run before letting an agent spawn hundreds of threads on your account. By BotBento Editorial · Published 2026-10-17 · Updated 2026-10-17 AI-assisted editorial: researched and drafted with AI using the primary sources linked below. Examples are illustrative; they are not measured product results. [Editorial policy](/editorial/). ## In this article - [What actually shipped on October 9](#what-shipped) - [A worked example: turning workflows on for a review agent](#worked-example) - [The limits, and what a finished run doesn't tell you](#limits-and-failure-paths) - [The decision: when this earns its complexity](#the-decision)[Read as Markdown ](/text/blog/claude-managed-agents-dynamic-workflows/index.md) ## Key takeaways - Dynamic workflows let the agent itself decide when to write and start a background program that fans a task out to many agents; you don't call an endpoint to trigger it. - A run is capped at 1,000 agent starts over its life and 64 threads working at once, and Anthropic says the 64 figure isn't guaranteed and can change. - A workflow run ending 'completed' does not mean the work succeeded; you still have to read each thread's events to find failed work. ## What actually shipped on October 9 Anthropic's Claude Platform release notes describe a new beta capability for Claude Managed Agents: dynamic workflows. The feature lets an agent write a workflow, defined as a program that runs many agents in phases and combines their results, and have the server execute it in the background rather than the primary agent delegating to each helper one request at a time. The mechanism replaces something narrower. Before this, a Managed Agents session could only hand off work through subagents, where the lead agent calls a helper itself and reads what it reports back one thread at a time. Dynamic workflows add a second path: the agent writes the orchestration logic, and that program runs many agent threads across phases without the lead agent staying in the loop for each one. Sources: [Claude Platform release notes - Claude Platform Docs](https://platform.claude.com/docs/en/release-notes/overview), [Multiagent orchestration - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration). ## A worked example: turning workflows on for a review agent Illustrative setup, based on Anthropic's own sample: an agent definition named Contract Reviewer sets model to claude-opus-5-5 and adds a multiagent block with type multiagent\_20261001 and workflows.type set to enabled. The system prompt tells the agent to start a workflow run when asked to review more than a few contracts, and to handle one or two itself without a run. Anthropic's docs note that under the multiagent\_20261001 type, dynamic workflows and subagent delegation are both on by default; if you want workflows only, you add a separate subagents.type: disabled entry to the same block. There is no separate call to start a run. You send a normal user message describing the batch of contracts, and the agent itself decides whether the task warrants a workflow, guided by whatever instruction you put in the system prompt. Once it starts one, the run executes as a sequence of named phases, each phase running one or more agents in their own session threads, and you follow progress through workflow\_run.\* events on the session's event stream rather than polling a status field. To register agents the workflow should reuse instead of letting it define everyone inline, you list them under workflows.predefined\_agents in the same multiagent block. That's also how you'd route part of the job to a cheaper or faster model than the one running the primary session. Sources: [Multiagent orchestration - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration). ## The limits, and what a finished run doesn't tell you Anthropic's workflow runs documentation sets a concrete ceiling on one run: 64 threads can work at once, and the run creates no more until one finishes. The docs are explicit that this number isn't a promise: the API doesn't guarantee it, so it can change. Separately, a workflow can start up to 1,000 agents across the whole life of one run; go past that and the run ends with a thread\_limit\_error rather than quietly degrading. The failure paths matter more than the headline limit. The documentation lists the possible outcomes of a run directly: a result of type completed means the workflow finished running, but the result doesn't say whether the work passed, and a run can end completed even though work on some of its threads failed or a thread couldn't even be created. Finding that out means reading the events of each of the run's own threads, not trusting the top-level status. Other documented outcomes include a stopped result when the agent or an archived session ends the run, a timeout\_error when the run hits its lifetime (24 hours by default, or shorter if the agent set one), a program\_error when the workflow's own code failed or broke a rule, and an unknown\_error that can also mean the run crossed one of the server's other, unlisted limits on workflows. That last point is worth sitting with before you rely on this in production: Anthropic's own docs say the server enforces additional limits and rules on workflows beyond the four it lists, without naming what they are. If a run fails for a reason outside the documented set, you'll see unknown\_error and have to debug from the thread-level events rather than from a clear error code. Sources: [Workflow runs - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/workflow-runs). ## The decision: when this earns its complexity Anthropic's own framing is useful here: dynamic workflows fit most work that needs more than one agent, specifically work with many pieces, parallel work, or a long task that should finish sooner, and the docs name audits, migrations, deep research, and cross-checking as examples. That's a narrower bar than 'anything involving more than one tool call.' A single support reply, a one-off calendar check, or a task that naturally runs as a short linear sequence doesn't need a program writing and coordinating other agents; a single agent with tools, or at most a couple of ordinary subagent delegations, covers it with far less to audit. The bar for turning this on for a small team should be: the task genuinely decomposes into many similar independent units of work (reviewing hundreds of documents, auditing many repositories, cross-checking many records), and you're willing to read thread-level events rather than trust a top-level completed status. If either condition isn't true, stick with subagents or a single agent, because every agent a workflow starts uses tokens independently, and a run has no separate price of its own on top of that usage. Before pointing this at anything that touches real systems, set a session budget at session creation (it can't be added later), write the system-prompt rule for when the agent should start a run at all, and build the check for 'completed but something inside failed' into whatever reads the run's result. Treat the beta header and the unspecified extra limits as a sign this is still settling: run it against disposable or sandboxed data until you've seen a few runs complete and inspected their thread events yourself. Sources: [Multiagent orchestration - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration), [Workflow runs - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/workflow-runs). ## Primary sources Sources checked 2026-10-17. Standards and product documentation can change; follow the linked version when implementing. - [Claude Platform release notes - Claude Platform Docs](https://platform.claude.com/docs/en/release-notes/overview) — Anthropic - [Multiagent orchestration - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration) — Anthropic - [Workflow runs - Claude Platform Docs](https://platform.claude.com/docs/en/managed-agents/workflow-runs) — Anthropic BotBento is in development. [Suggest a correction](/contact/).