If you only generate one candidate, you have no way to know if it was a good one. Fusion's answer is structural: run the same task through rival agents at the same time, in isolated lanes, and let the results compete.
How a tournament run works
- The task is frozen first: one spec, one Definition of Good, identical for every lane.
- Each candidate agent works in its own branch and worktree. No lane can see another lane's work.
- Every candidate runs the same objective gate: the tests, checks, and criteria the contract demands.
- A blind jury scores the survivors without knowing which vendor produced which candidate.
- An integrator assembles one final from the winner, grafting the best ideas from the losers where the judges approved them.
Why bother with parallel lanes
A single agent retrying the same task inherits its own blind spots on every retry. Different models fail differently: one is careless with edge cases, another over-engineers, a third misreads the spec in a way the other two do not. Running them against each other converts that diversity from noise into signal.
The dirty secret of "just regenerate it" workflows is that you, the human, become the tournament judge, one candidate at a time, with no criteria and no memory. Fusion moves that judging into the machine, against criteria that were frozen before anyone wrote a line.
Competition needs a referee
A race without a finish line is just traffic. The parts that make the tournament honest are the frozen contract, the shared gate, and judges who cannot see brand names. Take away any one of them and you are back to picking the output that sounds most confident.
That referee role is the product. The lanes are just how we make it fair.