Dynamic workflows in Claude Managed Agents entered public beta on October 9, 2026, announced by Anthropic through its developer account ClaudeDevs. It is a new form of multi-agent orchestration: the lead agent first writes a plan that spans several stages, hands each stage to a group of agents working in parallel, and then merges the results, instead of one agent running everything end to end.
How it works in practice
Developers set the multiagent type to multiagent_20261001 in the agent configuration and ask Claude to run a workflow; Claude then writes the plan and schedules the execution itself. The stated ceiling is up to 1,000 agents orchestrated in a single run. In the demo, a workflow used the agent toolset to search 300 contracts for change-of-control clauses — a task where each document is independent, can be processed in parallel, and needs one combined summary at the end, which is exactly the shape dynamic workflows are built for.
Why this update matters
Multi-agent orchestration in Claude Managed Agents previously required a separate research-preview application. Moving to public beta turns access from invitation-only into something any developer can configure. For teams doing bulk document processing, codebase-scale changes, or long checklist reviews, what is saved is not model calls but the engineering work of building scheduling, sharding, and result-merging infrastructure themselves.
The cost is stated plainly: tokens
Anthropic also warns that dynamic workflows are powerful but can consume a lot of tokens, and suggests starting with a well-scoped task before gradually raising complexity. The 1,000-agent ceiling is a capability limit, not a recommendation; the finer a task is split, the more repeated context reading costs. An onboarding path is provided: a managed-agents onboarding command in Claude Code, plus templates on the official quickstart page.
For teams deciding whether to build their own orchestration layer, the test is straightforward: if a task is naturally parallel and its results can be merged, run a small sample in the public beta first and compare the token bill against the single-agent version before scaling up.
