Symbiotic Pruning
Agents check each other's work. An agent that keeps failing peer checks is removed and its work is handed to healthy peers. A quorum rule stops one bad critic from pruning good agents.
How it works
Each round, every agent does a share of tasks. Pairs of agents re-run a sample of each other's tasks against a reference and vote pass or fail. An agent that fails K rounds in a row is pruned and its load is split across the survivors. With the quorum guard on, a failure only counts if at least two thirds of an agent's checkers agree. With it off, any single fail vote counts, so a dishonest critic can knock out good agents one by one.
Claimed by the proposal, not verified
- "Up to 40% reduction in aggregate compute footprint" versus fixed topologies. No setup, workload or baseline is given. In this toy, compute falls only by the share of agents pruned, and fewer agents means less spare capacity.
- "Semantic fidelity within 0.02% error margin." There is no stated way to measure semantic fidelity, so this number can't be checked.
- "Field deployments" are mentioned with no named system or source.
- Peer checks are not free. Re-running a sample of a peer's work adds compute that the 40% figure doesn't account for. This toy charges 10% overhead per round for checking.
What to try
- Turn the quorum guard off with one dishonest critic. Watch good agents get pruned and the error rate stay flat.
- Set K to 1 with the guard on. Unlucky good agents start getting pruned by noise.
- Set faulty agents to 0. The best case is that nothing gets pruned, so compute savings only exist when you already had broken agents.
- Raise dishonest critics to 3 with 6 agents. A small swarm can't outvote them.
What a real build would need
Reference specs to check outputs against, a sampling budget for re-running peers' tasks, signed votes so a critic can't vote twice, and a human sign-off before any agent is removed in a five-agent fleet, where losing one agent is 20% of capacity.