Anthropic adds workflows that let Claude divide tasks among teams of AI agents

Claude can now write a program that delegates a large task to other AI agents, rather than handle everything in one conversation. Anthropic has added dynamic workflows to Claude Managed Agents, which its documentation labels beta. The feature coordinates background work and combines results; Anthropic’s bug-finding test suggests a substantial gain over a single agent.

From one conversation to a running program

The starting problem is work too large for one conversation. Anthropic’s workflow documentation gives reviewing hundreds of documents as an example. Instead of keeping that whole assignment in a single exchange, the agent writes a workflow: a program that starts other agents, collects their answers and combines them.

A workflow run is one execution of that program. Anthropic’s server runs it in the background, creating separate conversation threads for the agents as needed. The main agent can keep working, end its turn or check progress while the run continues. A session can have several runs open at once.

The program can split an assignment into named phases. Agents can work concurrently, then pass their findings to another agent for reconciliation. It can also repeat a step, choose the next action from an agent’s answer, or handle a failed agent. The sequence is therefore more flexible than sending the same request to many agents.

The Decoder reports support for up to 1,000 agents per execution. Its October 9 account also describes Anthropic’s test: the company hid 70 bugs in a 116,000-line codebase. A single agent found between 14 and 27 bugs per run, while the dynamic workflow consistently found 66.

Agents share files, not a single conversation

Each worker has its own thread, but the threads share the session’s sandbox—the working environment containing its files. That includes files from a mounted memory store. Separate conversations therefore do not mean separate copies of the underlying files; agents in a run work with the same material.

A workflow can use predefined agents or define agents itself. An agent created by the workflow receives a system prompt written by that program, rather than the main agent’s prompt. It uses the main agent’s model. Its available tools, external tool servers and skills are a subset of the main agent’s, and tools retain their permission policies.

Developers follow progress through the session’s event stream, which carries run creation, phase changes and the final result. Workflow-run events do not trigger webhooks, the automatic notifications often used to alert another service. The interface also exposes each worker’s thread, so developers can inspect its events and answer tool calls.

Control works at the run level. Only the agent starts a run, and developers cannot stop an individual workflow thread by its ID or archive it while the run remains open. Anthropic directs users to ask the agent to stop the run. Archiving the session can also end it.

A completed run still needs checking

The result marked completed means the workflow finished running—not that its work passed. A run can receive that status even if work inside a thread failed or a thread could not be created. Anthropic tells developers to read the worker-thread events to find those failures, rather than rely on the final run status alone.

Spending and time impose different controls. Reaching the session’s budget can pause a run; raising or removing that budget can resume it. But a run’s default lifetime is 24 hours, and the clock continues during pauses or while waiting for the client. A paused run can therefore expire instead of continuing later.

To enable the feature, The Decoder says developers select the multiagent_20261001 agent type. Anthropic’s documentation also describes a workflows setting in the agent’s multiagent configuration that turns dynamic workflows on or off. Claude Code users can begin through the onboarding command below.

/claude-api managed-agents-onboard

Anthropic recommends starting small because workflows can consume many tokens, the units of text models process. The bug-finding result is a company-reported test on planted defects. Whether that improvement carries over to other tasks remains unresolved; the initial result is not a demonstrated gain for every workload.