Claude Managed Agents' dynamic workflows: can one agent really orchestrate a thousand?

What it is

Dynamic workflows is a beta feature inside Anthropic's Claude Managed Agents. Instead of you hand-wiring a multi-step orchestration, the lead agent writes a program itself and Anthropic's servers run it in the background. Anthropic put dynamic workflows for Claude Managed Agents into public beta on October 9, 2026: an agent writes a program that runs many agents in phases and combines their results, and the server runs it in the background.

The headline number: Anthropic's Claude can now orchestrate up to 1,000 AI agents in parallel through dynamic workflows. That's Anthropic's cap, not a benchmark.

The next-work-session test

Scenario: you're an SDK developer with a 400-document contract review task you'd normally shard across a Python script, a queue, and a fan-in step. With dynamic workflows, you call the Managed Agents API once with the task description and session budget. The lead agent writes the orchestration program, kicks off phased subagent runs, and streams a combined result back. You spend your session on the eval harness and the schema the agents return, not on glue code.

What changes in your loop: fewer orchestrator bugs you have to own, but a new class of "why did the lead agent write that program?" traces to review. See the update log on Releasebot for the dated beta entry.

Pricing

Managed Agents is billed on two axes — model tokens and session runtime — and workflow runs don't add a separate line item. The workflow runs page says the tokens its agents use "are billed like the session's other tokens, at each model's rates". Anthropic's pricing page says Claude Managed Agents is billed on two dimensions: tokens and session runtime.

Session-level cost control is in your hands, not the model's: A session budget caps that spending, and you can only attach one when you create the session.

For a concrete illustration (third-party, worked from Anthropic's docs rather than from me): Anthropic's own documentation walks through a one-hour coding session on Claude Opus 5 that consumes 50,000 input tokens and 15,000 output tokens: Input tokens | 50,000 × $5 / 1,000,000 | $0.25 Output tokens | 15,000 × $25 / 1,000,000 — plus runtime. Treat the per-agent cost of a 1,000-agent run as the dangerous unknown; model it before you ship.

What we'd actually use it for

Narrower than "autonomous research team": embarrassingly parallel passes over a known corpus. Classifying tickets. Extracting a fixed schema from a document pile. Running the same eval across N prompt variants. The lead agent writes the fan-out, you write the per-item prompt and the merge rule.

For a day-to-day loop, the realistic unit of work is dozens of subagents, not a thousand. Claude can now write a program that sends work to other agents, checks what comes back, and combines the results while the main conversation remains available. That "main conversation stays available" bit is the real developer-experience win — your REPL isn't blocked while 200 extractions run.

Limits

It's beta, and it's opinionated. subagents와 workflows는 기본 둘 다 켜지며, 워크플로만 쓰려면 subagents를 disabled로 둔다 — i.e., both subagents and workflows are on by default; if you want workflow-only behavior, you have to explicitly disable subagents when configuring the agent.

Other honest limits:

• The lead agent decides when to spawn a workflow. Claude Managed Agents can now write a program that runs up to 1,000 agents, and the model decides for itself when to set one going. That's great until it isn't — expect to add policy guards and session budgets as your first line of defense.
• Observability of a 1,000-agent run is still a hard problem the platform doesn't fully solve. You'll want structured logging on every subagent output.
• No separate SLA for the generated workflow program — if the lead agent writes a bad plan, you pay for it in tokens.
• Vendor lock-in is real: this orchestration lives inside Anthropic's managed runtime, not your Kubernetes cluster.

Try it if

• You already use Claude Managed Agents and have at least one task that fans out over a corpus.
• You have an eval harness that can score aggregated outputs, not just single calls.
• You're happy to pay for runtime as well as tokens in exchange for deleting orchestration code.
• You want to prototype a multi-agent pattern without standing up your own queue.

Skip it if

• Your agent loop is a single-turn tool-calling pattern — this is overkill.
• You need deterministic, auditable orchestration (regulated workflows, financial reconciliation). A plan written by a model at runtime is the wrong shape.
• You're multi-model by policy — this is Anthropic-only infrastructure.
• You can't yet put a hard session budget on a run you don't fully control. Fix that first, then come back.