The Pulse
Anthropic Adds Dynamic Workflows to Claude Managed Agents
Anthropic added dynamic workflows to Claude Managed Agents on October 9, 2026, allowing an agent to write a program that coordinates multiple agents in the background. The beta feature supports phased work, parallel agents and repeat rounds
AI.info Team · Written with AI ·
Claude can write the workflow, not just assign tasks
Anthropic added dynamic workflows to Claude Managed Agents on October 9, 2026, giving an agent the ability to write a program that coordinates multiple agents in the background. The company’s release note describes the feature as intended for work with many components, including reviewing hundreds of documents. The release is in beta and requires the managed-agents-2026-04-01 beta header.
Earlier multiagent orchestration let a session agent delegate tasks to other agents, but the agent itself determines when to delegate and what happens next. Dynamic workflows add a way for the agent to write a sequence of work for other agents, gather their results and use those results to determine what happens next. Anthropic’s orchestration guide describes agents working in parallel in separate session threads, each with its own conversation history.
Phases, parallel work and repeat rounds
A workflow can split work into phases, such as having several agents examine separate documents before another agent synthesizes what they found. The program can pass one agent’s result to another, repeat work within a phase, choose a next step based on a result or handle a failed agent. Agents in a run use the session’s sandbox and its shared files, while working in separate threads.
The session’s main agent starts the run, but the server executes it in the background. The main agent can keep working or end its turn while the workflow runs. When the run ends, the agent generally gets another turn to read the results. An interrupt can prevent that turn from arriving; after an archive or termination, it does not arrive. Anthropic says developers can follow progress through workflow_run.* events on the session event stream and inspect the run’s agent threads.
Beta setup puts the agent in charge of starting runs
To enable the feature, Anthropic’s release note directs developers to set the agent’s multiagent field to the multiagent_20261001 type and enable workflows. The agent determines whether and when to start a run in response to the work it receives; developers can guide that choice through the agent’s system prompt. The workflow-run documentation says only the session’s primary agent can start a run, so agents inside a run cannot launch nested workflows.
The beta also brings operational limits. A run has a default lifetime of 24 hours, and a session can have up to 10 open runs by default; idle runs count toward that limit. A run marked completed does not, by itself, establish that every piece of work succeeded, so developers must check its agent threads for failures, according to Anthropic’s workflow-run guide.
A new layer on May’s orchestration beta
Anthropic made multiagent orchestration available in public beta on May 6, 2026, months before adding dynamic workflows. The October feature lets the agent write a program to coordinate agents and pass results between them, while leaving developers responsible for configuring when runs should start and monitoring their execution.
For teams handling large batches of documents or other tasks that divide into stages, the practical change is that Claude Managed Agents can now assemble and run that staged process within a session. The feature remains beta, and its run status must be checked alongside the results from the individual agent threads.