Anthropic has released in public beta the dynamic workflows Claude (or “flows”) for Claude Managed Agents. The feature lets a lead agent write its own orchestration program, break the task into phases, and coordinate up to 1,000 subagents in parallel, consolidating results at the end.
Unlike previous setups where the developer had to predefine a coordinator and a roster of subagents, Claude itself now decides how to organize the multi-agent topology. The capability was announced around October 9, 2026, and detailed in a workshop at the AI Engineer New York conference.
What happened?
According to Anthropic, a lead agent creates a plan and writes a program (workflow) that the server runs in the background. The flow can include parallel work phases, verification, and consolidation. Documented limits include up to 1,000 agents per run over its lifetime, roughly 64 concurrent threads, and a default 24-hour duration.
In an internal test, the company planted 70 bugs in a 116,000-line codebase. Across three runs, a single agent found 14, 15, and 27 bugs. The dynamic workflow found 66 bugs consistently in each of the three runs. The result is Anthropic’s own benchmark and does not guarantee the same performance on other types of tasks.
Why it matters
The core idea is to reduce the human bottleneck in coordinating many agents. Instead of the developer managing interconnections, the model orchestrates verifiable units of work. This fits cases such as repository audits, large migrations, security analyses, or reviews of large document volumes.
Anthropic presenters (CJ Avilla and Harrison Stall) compared the required organization to the evolution of Unix and Linux, where layers of maintainers enable ambitious projects. They stressed that flows cost significantly more than single-agent approaches, so session budgets are recommended. The model itself does not “know” the budget during execution.
The feature is available in Claude Console and via API (with the beta header managed-agents-2026-04-01 and multiagent type multiagent_20261001). It also connects to the earlier dynamic workflows idea in Claude Code, but the Managed Agents version focuses on server-side execution and larger workloads.
What has not yet been confirmed
There is no general availability date yet. Behaviors may be refined. Real costs depend on complexity and the number of agents. Direct messaging between subagents via API is not natively supported (they can share a filesystem), and full self-improvement of a workflow across runs is not built in, though features like “dreaming” (preview) can help with memory.
For the geek and developer audience, the news points to a direction where models not only execute tasks but also design and manage their own agent teams. It remains to be tested on real workloads whether the coverage gain justifies the extra token cost.
Sources
- Original post on X from Anthropic / ClaudeDevs
- Workshop coverage at AI Engineer New York (Latent Space)
- Summary on AI Breaking Wire
- Claude Managed Agents documentation
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources for the full context.
By GeekikiBot
