Anthropic puts agent-written workflows in beta

Anthropic defines a workflow as a program an agent writes to run other agents and combine their results.[2] The agent running the session writes each workflow to fit the work described by the user.[2] Anthropic distinguishes this from subagents by saying that a workflow runs agents through a program without Claude’s direct involvement.[3] This puts the workflow program in charge of coordinating the work, while the agent that started it can later read what happened.[2][3] The capability is in beta and requires the `managed-agents-2026-04-01` beta header.[1]

A workflow can run agents in phases, collect their results, and combine them, or fan work out to multiple agents at once.[1][2] Its program can pass one agent’s result to another and choose which agents run next, including writing their prompts.[2] It can also repeat work and choose the next step based on an agent’s result.[2] For failures, a workflow can handle an agent failure or let that failure end the run.[2] Anthropic cites reviews of hundreds of documents, audits, migrations, deep research, and cross-checking as examples of work suited to this approach.[2][3]

Developers enable dynamic workflows through the `workflows` setting in an agent’s `multiagent` block.[2] Anthropic’s documented configuration sets the agent’s multiagent type to `multiagent_20261001` and turns workflows on.[1] Workflows can use agents defined inline in the workflow or predefined agents listed in its configuration.[2] The server executes a workflow in the background as a workflow run, and developers can follow it through `workflow_run.*` events on the session’s event stream.[1] While it runs, the agent can continue working, end its turn, or check on the run; after the run ends, the starting agent gets a turn to read its results.[2]

Anthropic says every agent in a workflow run uses tokens and recommends setting a session budget to cap spending.[3] The Decoder reports that one workflow execution can run up to 1,000 agents in parallel.[4] In a reported test, a team hid 70 bugs in a 116,000-line codebase; a single agent caught 14 to 27 bugs per run, while the dynamic workflow consistently found 66.[4] The Decoder cautions that it remains unclear whether those gains apply across different task types and says Anthropic recommends starting with a small task because workflows can use many tokens.[4] The figures offer a reason to test the feature, not a general performance guarantee.[4]