Claude agent writes its own orchestration code – Dynamic Workflows in beta costs $0.08 per hour
On October 9, 2026, Anthropic launched "Dynamic Workflows" in beta for Claude Managed Agents: a lead agent writes a program that lets it split large tasks across hundreds of sub-agents and gather the results. But the number making the most noise in coverage — up to 1,000 agents per run — is a cumulative limit, not parallel execution.
What's new
Claude Managed Agents is Anthropic's cloud-based agent-execution infrastructure. With the October 9, 2026 launch, an agent can now itself write a program that runs multiple agents in stages and gathers the results, executed in the background as a so-called "workflow run" (つくるAIラボ, citing Anthropic's release notes). The program-writing itself is the point: instead of a developer coding a fixed orchestration pipeline, the lead agent generates the work procedure dynamically based on the task it is given.
How it works
The flow is described in three steps (ぴすけ, citing Anthropic's announcement): the lead agent creates a plan and writes the procedure, many sub-agents work in stages, and the results are combined and returned to the lead agent. The Decoder summarizes the same pattern: a lead agent plans, distributes tasks, and merges the results when they are done (The Decoder).
In Anthropic's demo, 300 contracts were searched for "change of control" clauses, and 41 were identified (ぴすけ). This illustrates the work pattern: a review that would otherwise require either many hours of manual work or a slow, sequential agent run is parallelized across many sub-agents.
Activation requires no separate API call. You set multiagent.type in the agent definition to multiagent_20261001, and both workflows and delegation to sub-agents are then enabled by default (XenoSpectrum). Concretely, you set the agent's multiagent field to {"type": "multiagent_20261001", "workflows": {"type": "enabled"}} (つくるAIラボ). During execution, workflow_run.* events appear in the session's event stream, and the feature requires keeping the managed-agents-2026-04-01 beta header.
The limits behind the headline
This is where reality diverges from simple coverage. The Decoder wrote that up to 1,000 agents can run "in parallel" per run — but according to XenoSpectrum, based on Anthropic's workflow-run specifications as of October 10, 2026, 1,000 is a cumulative limit for the entire run: "a cumulative total across the whole run; it does not mean 1,000 agents run simultaneously." Actual concurrency is 64 threads per run — the next agent cannot start until a slot frees up. XenoSpectrum adds that this value can change and is not API-guaranteed (XenoSpectrum).
For users, that means throughput is limited by 64 concurrent threads, not 1,000 — a significant point for both latency and cost estimation.
What it costs
Pricing is pay-as-you-go: $0.08 per hour of runtime, plus tokens. At the same time, Anthropic warns that workflows can consume a large number of tokens and recommends starting with tasks of limited scope (つくるAIラボ, citing the release notes and Anthropic's official ClaudeDevs X account). The fact that the vendor itself recommends cautious scaling is worth noting: a multiplier principle in which one lead agent spawns dozens of sub-agents quickly produces token consumption that is hard to predict in advance.
The evidence of value — and its limits
Anthropic's own benchmark is so far the clearest sign that the orchestration actually helps: the team hid 70 bugs in a 116,000-line codebase. A single agent found between 14 and 27 per run, while the dynamic workflow consistently found 66 (The Decoder, which explicitly attributes the result to Anthropic's own testing).
Two caveats apply. First, these are company-driven figures, not independently verified. Second, The Decoder points out that it remains to be seen whether the gains hold across task types other than debugging code. The contract search in the demo points toward broader applicability, but it is a demonstration, not a benchmark.
Context and open questions
The name can cause confusion: Anthropic launched "Dynamic Workflows" for the developer tool Claude Code back on May 28. The October 9 announcement brings the same-named mechanism to Claude Managed Agents' cloud environment, but with a different execution environment and different limits — according to XenoSpectrum, the two should be treated as separate features (XenoSpectrum).
A fair amount remains to be clarified. The real-world token economics in practice — how $0.08 per hour relates to token spend in typical runs — is not yet documented in available reporting. Generalization of the benchmark result to other domains is unresolved. And terminology varies between "beta" in the release notes and "public beta" in the ClaudeDevs post; since both use the same beta header, some coverage treats them as equivalent, but that is an interpretation, not confirmed. As of today, the safe summary is: a promising orchestration mechanism in beta, with an impressive cumulative limit, a modest real parallelism of 64 threads — and a proof of value that so far comes mainly from the vendor itself.