Definition
Plain language
When splitting a task between multiple AI agents makes them do worse than a single agent doing it alone.
As stated in the literature
An observed parallelism penalty in multi-agent settings where two cooperating agents on a shared task achieve a lower joint success rate than a single sequential agent, attributed to coordination overhead.
Why it matters: It warns that adding more agents isn't automatically better, since the overhead of working together can outweigh the benefit of dividing the work.
For example, two agents splitting a coding task can finish less successfully than a single agent doing the whole thing alone, because they spend effort syncing up.
Heard on the show
“That result has a name — the curse of coordination — and the team behind today's paper had a fix.”Episode 096 — How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty