Claude Managed Agents Adds Dynamic Workflows: Parallel Agents, Phases and a Shared Budget

Anthropic’s [October 9 Claude Platform release notes](https://platform.claude.com/docs/en/release-notes/overview) introduce **dynamic workflows** for Claude Managed Agents. The agent can write a program that launches other agents in phases, then combines their outputs. The server runs that program in the background as a workflow run. This extends the Managed Agents platform; it is **not** a new Claude model or a general Claude Code update.

What can developers do with it?

Imagine reviewing hundreds of files: one phase distributes files to specialist agents, while a later phase checks and combines their findings. Anthropic cites document review as an example. The agent decides whether to start a run based on the task and its instructions; the developer does not invoke a separate “start workflow” API call. Progress and errors appear as `workflow_run.*` events on the session’s event stream. [Sources: Anthropic’s orchestration guide](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration); [workflow-run documentation](https://platform.claude.com/docs/en/managed-agents/workflow-runs)

To enable the beta, use the `managed-agents-2026-04-01` header and set the agent’s `multiagent` configuration to `{"type":"multiagent_20261001","workflows":{"type":"enabled"}}`. Anthropic says workflows and subagents are both enabled by default with that multiagent type, so teams wanting **workflows only** should explicitly disable subagents. A system-prompt instruction can tell the agent when a run is worth starting and when to handle a smaller task itself. [Sources: Anthropic release notes](https://platform.claude.com/docs/en/release-notes/overview); [configuration guide](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration)

Cost and operational limits

A workflow run has **no separate fee**. Its agents’ tokens are billed at each model’s rates and count against the session budget. Claude Managed Agents also charges **$0.08 per running session-hour**; applicable tools can add charges. Anthropic recommends setting a session budget because parallel agents can consume tokens concurrently. [Sources: workflow-run documentation](https://platform.claude.com/docs/en/managed-agents/workflow-runs); [Managed Agents pricing](https://platform.claude.com/docs/en/about-claude/pricing)

The current documentation lists **64 threads working at once per run** and **10 open runs per session by default**; Anthropic says the concurrency number is not guaranteed to remain fixed. A workflow can start up to 1,000 agents over its lifetime. A failed agent may be started again on a new thread, so write side-effecting tools to tolerate duplicate calls. [Source: workflow-run limits](https://platform.claude.com/docs/en/managed-agents/workflow-runs)

**Bottom line:** dynamic workflows make sense when a task can be split into meaningful pieces and recombined. For a small coding request, the extra orchestration and token spend may not help. For larger audits, migrations or cross-checks, budget the session, watch run events, and verify the combined result before acting on it.