Anthropic opens dynamic workflows in Claude Managed Agents as a public beta

Anthropic added dynamic workflows to Claude Managed Agents as a public beta on October 9, 2026. Managed Agents is the service where companies run Claude-based agents on Anthropic's servers, with a ready-made environment and tools. With the new feature, the agent running a session writes a small program that breaks a large job into pieces. The program hands those pieces to many agents in phases, collects what each one returns and combines the results at the end.

According to Anthropic's documentation, the feature is meant for work with too many pieces for a single conversation: reviews of hundreds of documents, audits, migrations, deep research and cross-checking are the examples given. The program can run many agents at once, pass one agent's result to another and have a draft revised until a review passes. The agents in a run share the session's files. A similar feature launched in Claude Code earlier this year; there the program runs on the user's machine, while in Managed Agents it runs in the background on Anthropic's servers.

How to turn it on

The feature uses the managed-agents-2026-04-01 beta header. To enable it, the agent's multiagent field is set to the multiagent_20261001 type with workflows enabled. Developers tell the agent in its system prompt when to start a run, and only the agent starts one. Progress can be followed phase by phase through workflow_run events on the session's event stream. A workflow can define its own agents or pick from agents the developer has set up in advance. ClaudeDevs, Anthropic's developer account, said developers can start from a ready-made bug hunter example in Claude Code with the command /claude-api managed-agents-onboard bug-hunter.

Limits and spending controls

According to the documentation, a workflow can start up to 1,000 agents over the life of a run. There are also limits on how many agents work at once, how long a run lasts and how many runs a session can keep open; the current values are in the docs. A run has no price of its own: the tokens its agents use are billed like the session's other usage, at each model's rates. Anthropic recommends setting a session budget to cap spending. When the budget is reached, open runs pause, and they resume once the budget is raised.

In a test shared by ClaudeDevs, Anthropic planted 70 bugs in a 116,000-line codebase. A single agent found 14, 15 and 27 bugs across three runs, while the dynamic workflow found 66 in each of its three runs.

The feature is in beta, and because every agent in a run uses its own tokens, costs can grow quickly; Anthropic suggests starting with a scoped task and raising the complexity step by step. The bug-finding test was run by Anthropic. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Claude Platform release notes, Claude docs: Multiagent orchestration, Claude docs: Workflow runs, ClaudeDevs announcement.