Reviewing hundreds of documents need not mean asking one AI agent to read them in sequence. Claude’s new dynamic workflows beta lets an agent write a program that distributes the work, runs it in phases and brings the findings together.
Based on 7 reviewed sources · See sources
Claude writes the coordination plan
The 9 October release note adds the feature to Claude Managed Agents, Anthropic’s developer service for running agents. Its multiagent documentation explains the distinction: instead of the lead agent supervising every hand-off itself, a program coordinates the work in the background. The lead agent can remain available to the user.
Developers enable dynamic workflows in the agent’s configuration and describe the task. Anthropic’s original announcement introduces the public beta and shows the approach in action.
Anthropic announces a public beta for dynamic workflows in Claude Managed Agents: a lead agent plans phased work across multiple agents and combines their results.
View the original post on X ↗1,000 agents does not mean 1,000 at once
The workflow-run documentation separates the lifetime allowance from simultaneous work. The advertised 1,000 agents can be spread across successive stages; the current limit is 64 active agent threads in one run. A thread is an agent’s separate conversation.
The beta’s operating limits
Per run, except the session-wide open-run allowance
| Limit | Documented value |
|---|---|
| Agents started over a run’s lifetime | 1,000 |
| Agent threads working simultaneously in one run | 64 — subject to change |
| Run lifetime | 24 hours by default; the agent can set it |
| Open runs in one session | 10 by default, including paused runs |
The coordination is easier. The bill still needs watching.
There is no separate workflow-run fee, but that does not make a run free. The pricing page bills model tokens at each model’s rates, plus $0.08 per session-hour while the session is running. A session budget pauses work when reached, although model requests already underway can take spending past it.
Anthropic’s developer account itself recommends starting with a small task because workflows can consume many tokens. Christina @ATX, whose profile describes her as an agentic systems engineer, asks in a reply whether efficiency has improved over her earlier workflow experiments. She supplies no matched cost logs or identified test of this beta.
The useful change is having Claude organise a large task as well as perform its pieces. Whether that produces better results for the money remains a question for controlled tests, rather than a conclusion to draw from the agent count.
Sources
- Dynamic workflows beta release note
Claude Platform Docs · Announcement ·
- Multiagent orchestration
Claude Platform Docs · Documentation
- Workflow runs, budgets and limits
Claude Platform Docs · Documentation
- Claude Managed Agents pricing
Claude Platform Docs · Documentation
- Dynamic workflows public-beta announcement
ClaudeDevs on X · Announcement ·
- Token-use caution in the announcement thread
ClaudeDevs on X · Announcement ·
- A practitioner's question about workflow efficiency
Christina @ATX on X · Research ·
Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.
