What if the most instructive document in AI agent design right now isn’t a paper from a frontier lab, but the fight record of a bantamweight ranked #179 in the world?
Malcolm Wellmaker is scheduled to face Otari Tanzilovi on October 10, 2026, at UFC Fight Night: Allen vs. Duncan in Las Vegas. He arrives there carrying two Performance of the Night bonuses, roughly $100,000 in reward signal for doing one thing extremely well, and two consecutive losses that suggest the thing he does extremely well is no longer sufficient. I study agent architectures for a living, and I have rarely seen a cleaner natural experiment in what happens when a high-variance policy meets a reward function that only pays out on the tail.
A policy tuned for the tail
Consider what the bonus structure actually rewards. Performance of the Night is not awarded for winning. It is awarded for a particular kind of winning: fast, violent, legible. Wellmaker, a Dana White’s Contender Series alumnus from the 2024 season, collected that bonus twice, most recently at UFC on ESPN 69 — Usman vs. Buckley in 2025. Two large payouts for a specific behavioral mode.
Any reinforcement learner would read that gradient the same way. The aggressive, power-forward policy got reinforced hard. And because the bonus is a tail event, the policy that earns it is necessarily a high-variance one. You cannot reliably produce a knockout the way you can reliably produce a decision win. The strategy that generates spectacular outcomes also generates catastrophic ones, drawn from the same distribution.
The record bears this out. On November 15, 2025, Wellmaker lost a unanimous decision to E. Ewing across three rounds. On May 16, 2026, at UFC Fight Night: Allen vs. Costa, he lost by submission at 4:08 of the second round to J. Diaz. A decision loss and a submission loss are two different failure modes, and both are what the long tail looks like when it points the wrong way.
What a correct credit assignment sounds like
Here is what interests me most as an architecture person. Wellmaker’s own diagnosis is more precise than most post-mortems I read in engineering retrospectives: “I need to go in there, still be aggressive, still try to put this power on people, knock them out cold, but, you know, just do it in a more strategic manner.”
Read that as a policy update. He is not discarding the behavior that earned the reward. He is not flipping to a conservative, risk-minimizing mode, which is the standard overcorrection after two bad episodes. He is proposing to keep the action space intact while improving the selection function over it. In agent terms: same primitives, better planner.
That distinction matters enormously and almost nobody gets it right. When agents fail, the common fix is to clamp the behavior. Lower the temperature, narrow the tool permissions, add guardrails until the system stops doing anything interesting. You get a system that fails less and accomplishes less. The harder fix, the one Wellmaker is describing, is to leave the capability in place and improve the conditions under which it fires.
Rankings as a lagging indicator
Fight Matrix puts him at #179. That number is a legitimate measure of something, and it is also a lagging indicator computed from past episodes. It tells you where the policy has been, not where it is heading. He remains on the UFC roster with a scheduled bout, which is the only forward-looking signal that actually binds.
We make this mistake constantly with agent benchmarks. We score systems on past performance distributions and then reason about them as though the score were a property of the system rather than a summary of its history. An agent mid-revision is badly described by its benchmark number. So is a fighter mid-revision.
His framing is explicitly about trajectory: “I’m excited to show that I’m in a place where I’m still growing, I’m still getting better, and my future is still bright.” You can dismiss that as athlete boilerplate. I would rather treat it as the correct epistemic stance for any learning system that has just received two strong negative signals and has not yet had the chance to act on them.
The design question
If you build agents, the question October 10 poses is worth sitting with. When your system earns outsized rewards for a specific high-variance behavior, and then that behavior starts producing losses, do you suppress it or do you teach the system when to use it?
Suppression is faster. Selection is better. We will get a data point in Las Vegas.
🕒 Published: