Claude Managed Agents dynamic workflows: background multi-agent runs for document-scale work
On October 9, 2026 Anthropic opened a public beta for dynamic workflows on Claude Managed Agents. A lead agent can write a program that fans work across phased agents in the background. This guide covers enablement, limits, budgets, and how enterprises should adopt it without exploding spend.
Jigar JoshiAgentic AI Architect and ConsultantIn this post (8 sections)
Introduction
Subagents already let a Managed Agents session hand work to specialists and wait for their replies. Dynamic workflows change the control plane: the lead agent writes a program that schedules phases, fans out threads, combines results, and can keep going while the primary conversation stays available. That is the Managed Agents counterpart to Claude Code graph-style orchestration, aimed at audits, migrations, and document-scale research.
Primary sources: Claude Platform release notes, multiagent orchestration, and workflow runs. Adjacent: Claude Code graph engineering and coordinator platforms.
What shipped
| Item | Value |
|---|---|
| Status | Public beta (Oct 9, 2026) |
| Beta header | managed-agents-2026-04-01 |
| multiagent type | multiagent_20261001 |
| Enable flag | workflows: { type: "enabled" } |
| Observability | workflow_run.* events on the session stream |
| Networking note | Limited networking also applies allowed_hosts to web_search and web_fetch (Oct 7 hardening) |
With `multiagent_20261001`, dynamic workflows and subagent delegation are both on by default. Teams that want workflows only can disable subagents in the same block. Only the primary session agent may start a run; agents inside a run cannot nest another run.
Limits that change capacity planning
| Limit | Documented value | Failure mode |
|---|---|---|
| Threads working at once in one run | 64 (subject to change) | Server waits until a thread finishes before starting more |
| Agents started over a run lifetime | 1,000 | thread_limit_error ends the run |
| Run lifetime | 24 hours default (agent may set shorter) | timeout_error; paused time still counts |
| Pricing | No separate workflow fee | Model tokens + session-hour charges; budget pauses runs |
How it compares with Claude Code workflows
Claude Code dynamic workflows keep orchestration in a JavaScript script the coding agent writes. Managed Agents dynamic workflows keep orchestration on Anthropic’s managed runtime with session streams, sandboxes, and MCP already wired. Use Code when the artifact is a repo change. Use Managed Agents when the artifact is a long research or ops job that should not occupy an IDE session.
What this means for developers
- Add the beta header and `multiagent_20261001` only in a staging project first.
- Put start/stop rules in the system prompt: when to open a run, when to stop, and what to return to the user.
- Subscribe to `workflow_run.created`, phase events, and `workflow_run.status_ended` before scaling concurrency.
- Pair with Haiku 5.5 workers and Opus or Sonnet 5.5 planners so fan-out cost stays predictable.
- Re-check limited-networking hosts: web_search and web_fetch now honor `allowed_hosts`.
What this means for businesses
- Document-scale agent jobs become a product feature, not a custom orchestrator project.
- Spend risk moves from “one long chat” to “hundreds of short agents.” FinOps needs session budgets on day one.
- Google’s Gemini agent (Oct 8) and Copilot fleets are competing coordinator surfaces. Pick one pilot workflow per platform.
- Security still owns egress: MCP, web_fetch, and sandbox networking do not disappear because work runs in the background.
Adoption checklist
- 01Pick one multi-document workflowContract review, policy audit, or migration inventory with a clear done definition.
- 02Enable workflows in stagingBeta header, multiagent type, session budget, and predefined agents list if used.
- 03Instrument eventsLog workflow_run IDs, phase progress, and ended result types including thread_limit_error.
- 04Cap concurrencyStart far below 64 concurrent threads. Raise only after quality and spend hold.
- 05Compare with Code and Gemini agentSame workload on Claude Code workflows and Gemini agent if the org already pays for those.
Conclusion
Dynamic workflows make Managed Agents a serious multi-agent runtime for work that is too large for one conversation. The beta is usable now if budgets, prompts, and event handling are in place. Without those, it is a cost amplifier dressed as automation.
Sources: Claude Platform docs ; Claude Platform docs ; Claude Platform docs ; Google blog
Agentic AI patterns, delivered Thursdays
What I am shipping, watching, and pruning out of client stacks each week. One email. No fluff.