Multiagent v2 playbook calls for 64 agents, diverse proof routes and adversarial checks

danshipper · x · 2026-07-21

A screenshot of multiagent v2 search-and-coordination guidance for theorem proving/ counterexample search.

It suggests using up to 64 concurrent agents and avoiding fixed allocations like “N agents for strategy X.”
Early rounds should maintain a genuinely diverse portfolio of proof and counterexample approaches, spanning areas like algebraic geometry, commutative algebra, topology, model theory, and computational sanity checks.
The document emphasizes keeping agents independent at first, tracking approach families explicitly, and redirecting agents when too many converge on the same incomplete route.
It also recommends adversarial checking: every proposed proof or counterexample should be tested for hidden assumptions, inversions, trivializing transformations, and incomplete noninvertibility arguments.
The desired outputs are concrete lemmas, constructions, equations, degree estimates, maps, or counterexamples—not vague optimism or “routine” status updates.

More from coding & agent