There’s something primal about watching two AI models fight for coding supremacy. The terminal blinks, tokens cascade down the screen, and somewhere in a GitHub repo, one model emerges victorious—while the other quietly dies of a ‘timeout’ error.
That’s the pitch behind Agon, a ‘coliseum’ for AI coding models. It’s a simple, addictive idea: put two models in an arena, give them a task, and let the better one win. The README practically sells itself. But after watching a few rounds, you start to feel the same itch one of Agon’s own commenters felt: ‘I’d rather there also be a collaboration between the models. Competition can only get you so far.’
The coliseum measures dominance, not capability. Those are not the same thing.
That comment is the whole argument in one sentence. Each model has a unique set of strengths and weaknesses. Collaboration allows those strengths to be pooled, while the weaknesses are hedged against broader perspectives. Instead of asking ‘which AI is best?’ we should be asking ‘how can these AIs make each other less bad?’
Agon isn’t wrong to create a competitive arena. It’s incomplete. By building a zero-sum game, we train ourselves to think of AI as a winner-take-all contest. That mentality seeps into the tools we build, the benchmarks we set, and the way we structure our development teams. We end up with one dominant model, one way of thinking, one point of failure.
But here’s the twist. The real future of AI-assisted coding isn’t a single supermodel. It’s a team of specialists that can challenge, verify, and improve each other’s work. A linter catches one kind of mistake. A test suite catches another. A second AI model can catch the ones that slip past both. That’s not a coliseum. That’s a band.
The sloppy terminal title in Agon’s interface—’peak sloppery’ one commenter called it—is actually a fitting metaphor. The whole project is raw, early, and unfinished. It’s a hackathon-worthy idea that hasn’t yet asked the harder question. And that’s fine. The question is still open.
So what should Agon become? Not a coliseum, but a negotiation table. Instead of pitting model A against model B, let them swap code, debate approaches, and vote on which solution survives. Let a weaker model route to a stronger one when it gets stuck. Let three different models generate three different answers, then run them all against a battery of tests to see which one actually works in practice.
The future isn’t one model that does everything. It’s many models that argue about the best answer—until they agree on the right one. That’s not a competition. It’s the opposite. And it’s exactly what we need to build next.
A coliseum crowns a champion. A lab makes a breakthrough. You get to choose where you work.
FAQ
Q: Isn't competition necessary to find the best model?
A: To a point. But benchmarks like these optimize for a narrow skill: winning a one-on-one match. Real-world coding is a team sport. You need a model that can read bad code, refactor it, write tests, and cooperate with a human. That's not a sprint. That's a relay.
Q: How would collaboration between coding models actually work?
A: Simple. Give them distinct roles—one writes, one reviews, one tests. Have them pass code back and forth like a PR workflow. If a model gets stuck, it can hand off to a specialist. The result isn't one genius. It's a well-managed team.
Q: What's the contrarian take?
A: The obsession with ranking AI models is holding back the field. Instead of asking 'who wins?', ask 'what can we build together?'. The strongest AI system won't be a single model. It'll be an ensemble of models that disagree just enough to catch each other's blind spots.