The Pulse
Anthropic Launches Managed Agent Workflows With a 1,000-Agent Cap
Anthropic launched dynamic workflows in beta for Claude Managed Agents on October 9, 2026. Each run can start up to 1,000 agents over its lifetime, but only 64 can work at once under the documented limit.
AI.info Team · Written with AI ·
Anthropic launched dynamic workflows for Claude Managed Agents on October 9, adding a way for an agent to divide a large task among as many as 1,000 agents over the life of a single run. The feature is in beta, and the limit is not a promise that all 1,000 agents work at once: Anthropic’s documentation sets a separate limit of 64 concurrent agents per run.
One agent writes the workflow
Dynamic workflows let the agent handling a session write a program for a task, start other agents to do pieces of it, gather their results and combine them. The program can split work into phases, pass one agent’s results to another, repeat a step, or choose what to do next based on an agent’s output. Anthropic gives reviewing hundreds of documents as an example of work that could exceed one conversation.
Each worker operates in its own session thread, using a prompt written by the workflow. It can be an inline agent defined for that workflow or a predefined agent configured in advance. The workflow runs in the background; its progress and status appear as events on the main session’s event stream. Users do not start runs directly: the agent on the session’s primary thread does.
The 1,000-agent ceiling is cumulative
The 1,000 limit counts agents started across the entire run, not agents active simultaneously. At the documented concurrent limit of 64, a workflow must finish work in some threads before creating more. Anthropic says the API does not guarantee that concurrency figure, so it could change.
The ceiling applies to starts, not necessarily to the total number of threads: the server can retry a failed agent in a new thread, meaning a run may create more than 1,000 threads. If a workflow requests another agent after reaching its cap, the server refuses to start it and ends the run with a thread_limit_error. Anthropic also documents a default run lifetime of 24 hours, which an agent can set shorter.
Beta setup brings budget and rate limits
Developers must enable workflows in an agent’s multiagent configuration and use the managed-agents-2026-04-01 beta header. Anthropic’s release notes specify a multiagent_20261001 configuration type and say developers can guide the agent through its system prompt, including when to start a run. Run progress is tracked through workflow_run.* events.
The feature does not have a separate run price. Agents’ token use is billed at the selected models’ rates and counts toward the session’s budget; reaching that budget pauses open runs until it is raised or removed. Runs also count against the organization’s existing rate limits. Anthropic’s documented default allows 10 open runs per session, with idle runs included in that total.
More parallel work, with a fixed execution budget
The new workflow model gives a single Claude Managed Agent a way to coordinate staged, multi-agent work without requiring developers to manage each worker as a separate top-level request. But the 1,000 figure describes the maximum number of starts across a run, while concurrency, elapsed time, session budgets and organizational rate limits constrain execution. Those controls make the advertised scale a ceiling for orchestration, not a guarantee of 1,000 agents working in parallel.