Anthropic updated its Claude Platform release notes on October 9 to say that dynamic workflows have entered Claude Managed Agents in beta. When a managed agent receives a large task, it can now write a workflow of its own: a lead agent breaks the job down, hands pieces to a group of subagents that work in stages, and then merges the results. A single run can coordinate up to 1,000 agents in parallel.
The setup is described in the documentation. A request needs to carry the managed-agents-2026-04-01 beta header, set the agent's multiagent field to the multiagent_20261001 type, and turn workflows on. The workflow runs in the background on the server, and the caller follows its progress through workflow_run events in the session event stream.
The 70-bug test
Anthropic's example is code review. The team planted 70 bugs in a 116,000-line codebase and asked a single agent to find them. Each run turned up between 14 and 27, with considerable swings from one run to the next. With a dynamic workflow, the result held steady at 66.
The figures come from Anthropic's own testing, and the task is of the kind that splits easily: a codebase can be cut into modules, each subagent checks one piece, and the pieces depend on each other very little. For work that needs continuous reasoning from start to finish, such as the argument chain of a long report, Anthropic has not published data on how much of the gain survives the split. The tech outlet The Decoder also cautioned that whether the result generalizes is something users will have to check against their own workloads.
Cost is the other open question. Anthropic warns in its documentation that these workflows can consume a large number of tokens and recommends starting small. The Decoder's report cites a senior OpenAI engineer's verdict on agent swarms:
"a massive waste of tokens"
The release notes do not list beta billing details or the range of tokens a single run is likely to consume.
One feature, two ways to run it
Dynamic workflows have already shipped in Claude Code, where they launched as a research preview in late May this year. Comparing the two versions, the differences come down to three points.
- Where the script runs: in Claude Code, the workflow script runs on the user's own machine; in Managed Agents, it runs on Anthropic's servers.
- Who starts it: Claude Code users can open the script, edit it and run it again; in Managed Agents, the agent itself decides when to launch a workflow.
- How costs are managed: according to a comparison from third-party developers, the managed version bills per session and lets you set a cap on each session.
Claude Managed Agents is a hosted service Anthropic launched in April this year for production agents that need to run continuously for hours or longer, with Anthropic responsible for the runtime environment. With dynamic workflows now available there, teams that build Claude agents into their own products do not need to assemble their own orchestration and concurrency layer: the lead agent splitting the task, the subagents executing it and the results being gathered all happen on the server side.
The Claude Code version is more of a developer's hand tool, letting you see every step and change the script at any point. The managed version is closer to handing the job off and waiting for the result. The first offers more control; the second moves operations onto Anthropic, and also passes part of the decision over how many agents a task spawns and how many tokens it burns to the model itself. The point enterprise customers care about most, the per-session spending cap, will have to wait for the general-availability documentation to show how it is set and what the default is.
The feature is still in beta, and interface parameters and field names may change later.
Sources: Claude Platform release notes, The Decoder, CocoLoop, ITHome; used to check the agent concurrency limit, the size of the test codebase and the number of bugs found.