A swarm of subagents is not throughput
Coordinators that fan out to thousands of subagents measure merged pull requests, which is output. The limit on fan-out is how much a team can check well enough that what lands is what customers need.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 25 September 2026 Updated: 1 min read
A coordinator can now run thousands of subagents. Ours can’t nest them past one layer.
Why our harness stops at one layer
A coding tool launched that coordinator this month. Our harness stops delegation one layer below the first delegate, and a quiet agent gets asked for its report, never replaced.
The reason is checking. Every delegate ends in a report that someone has to verify before anything ships.
One person can check a few reports between other tasks. A few dozen become a queue that grows faster than anyone works through it.
We have paid for fan-out going wrong, too. A delegate woken at the wrong moment re-sent its whole context, more than once in one evening.
Merged pull requests are output
The launch measures merged pull requests. That is output, and it says nothing about whether a customer got anything.
A change counts once it runs in production and customers use it, pay for it or tell you it helped. Faster only helps when it is the right thing.
Size fan-out by what your team can check
So the limit on fan-out is how much your team can check well enough that what lands is what customers need. A model release does not move that number.
Count how many results your team can check in a day, set your fan-out to that, and judge the result in production. Raise it only when checking gets better: stronger tests, hooks and gates that fail on their own, a first review by an agent.
(Source: Cursor, Introducing Projects, 10 September 2026)
Further reading
Sources and limits
-
Introducing Projects (opens in a new tab)
Method and limits
- Method
- Vendor launch post by Cursor describing a coordinator agent that plans work and delegates it to thousands of subagents, with a usage figure on merged pull requests.
- Limits
- Vendor announcement of its own product, and the merged pull request figures give no baseline, sample size or time window.
2026-09-10 | Verified: 2026-10-25
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Coding agent setup
- Generic coding agents hit a limit, custom agents go past it Claude Code, Cursor and Gemini CLI are built for the average codebase. Teams that scale with agents move from better agents to more agents to custom agents built for their own domain.
- An agent is a loop, and the loop is what you debug A coding agent perceives, decides, acts and observes, then repeats. Naming which of the four steps failed turns a mystery into a fix.
- Skills hold what agents need to know, sub-agents do the thinking The first architecture decision in an agent setup is sorting what needs to think from what just needs to know. Skills carry static knowledge, sub-agents do dynamic work, and the split decides the cost.
Scaling needs named owners
Coding agents scale across teams when named people own the platform, the team practice, the guardrails, and the measurement. Rollout & scale is where that ownership is built and handed to internal owners, in your stack.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check