Anthropic Adds Dynamic Workflows to Claude Managed Agents: 1,000 Parallel Sub-Agents, and 66 of 70 Planted Bugs Found

Anthropic has shipped the piece its agent platform was missing: orchestration. The company added dynamic workflows to Claude Managed Agents, allowing a lead agent to plan a task, distribute it to as many as 1,000 sub-agents running in parallel within a single execution, and merge their outputs when they finish. The feature is enabled by selecting the agent type multiagent_20261001, and developers can onboard through Anthropic's documentation or by running /claude-api managed-agents-onboard in Claude Code.

The headline result comes from Anthropic's own testing. The company hid 70 bugs in a 116,000-line codebase and compared approaches: a single agent found between 14 and 27 bugs per run, while the dynamic workflow consistently found 66. Those figures are Anthropic's claims, generated on a bug-finding benchmark the company designed itself, and they have not been independently verified. It is also unclear whether the gains transfer beyond this task type — bug hunting splits cleanly across code regions, while tasks with tightly coupled steps or shared state are less likely to see similar benefits, and merge errors can eat into them.

The infrastructure underneath is not new. Claude Managed Agents, which reached general availability earlier this fall alongside integrations with partners including Notion, Rakuten and Asana, already handled agent lifecycle, storage and tool access. Dynamic workflows add the multi-agent layer on top: until now, teams that wanted swarm behavior assembled it themselves with frameworks and custom glue code. Anthropic is effectively absorbing a layer of the LangGraph-and-scaffolding stack into its own platform.

The cost question is the open one. Anthropic acknowledges these workflows can consume "a lot of tokens" and recommends starting with small runs. No pricing for dynamic workflows has been disclosed, no per-run token figures were published, and it is not confirmed which Claude models power the lead agent versus the sub-agents, or what rate limits and concurrency quotas apply by plan. Standard model token rates would presumably apply to each sub-agent's usage, but the company has not said so explicitly.

The skepticism is already on record. A senior OpenAI engineer recently described agent swarms as a large waste of tokens. Anthropic's benchmark answers a different question than the one that determines whether swarms pay off: it measures detection rate, not cost per correct finding. A run that catches 66 of 70 bugs beats a single agent's 14 to 27 only if the token bill for the extra 39 findings is justified by their value — a number neither Anthropic nor its critics has published.

The strategic reading is more straightforward. Platform vendors are climbing the stack, from renting model calls to selling built-in orchestration, and the differentiation is shifting from raw model quality to how much of the agent framework a customer no longer has to build. Anthropic's own advice is the sensible interim stance: run small pilots on your workloads, and track tokens against outcomes before assuming a thousand agents is better than one.