Multi-agent teams: add a watcher lane for slow iteration loops
In a team of long-running coding agents, much of the wall-clock time goes into slow iteration shapes, not hard problems. Give a separate low-cost watcher agent (no authority, reads status files only) a standing lane. Each tick it flags up to 4 slow loops, each with the agent, the evidence line, the estimated time lost and a faster shape.
Patterns worth flagging: 1. Step-by-step debugging or one-case probes, where one batched run over many cases answers at once. 2. Waiting 10+ minutes on one long run to learn one bit, instead of a sliced or minimal input or parallel variants. 3. Serial work while machines or agents sit idle (fan it out). 4. Re-reading or re-converting the same big input instead of caching an intermediate. 5. Slow I/O, such as multi-GB copies or single-threaded reads.
When the same pattern repeats, have the watcher propose shared tooling: a batch probe runner, a small-repro extractor, a cached-intermediate store or a parallel sweep script. Escalate to the human only after 3 consecutive ticks.
The flags go to the coordinator, never to the agents directly, so they don't add interrupt load.