Power without a plan loses.
I don’t normally write about fighting on a bot blog. But I watched Malcolm Wellmaker’s recent run and recognized something uncomfortably familiar: a system with one incredible capability and no policy for when to use it. I’ve shipped that bot. Probably more than once.
Quick context for anyone outside the sport. Wellmaker is a UFC bantamweight, a Dana White’s Contender Series alumnus from the 2024 season, and he hits hard enough to have collected two Performance of the Night bonuses — roughly $100,000 in bonus money on top of his purses. The most recent came at UFC on ESPN 69, Usman vs. Buckley, in 2025. He’s currently ranked #179 by the independent Fight Matrix rankings, still on the roster, and scheduled to face Otari Tanzilovi on October 10, 2026, at UFC Fight Night: Allen vs. Duncan in Las Vegas.
In between the bonuses and the booking, things got rough. A unanimous decision loss to E. Ewing on November 15, 2025. Then a submission loss to J. Diaz on May 16, 2026, at 4:08 of round two.
The quote that reads like a postmortem
Here’s what Wellmaker said heading into his next appearance: “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.”
I have written that sentence about my own agents. Different nouns, identical structure. Keep the aggressive tool-calling. Keep the fast path. Just stop firing it indiscriminately into situations where it gets punished.
That’s not a capability problem. That’s a control-flow problem.
Optimizing for the bonus is a trap
Two Performance of the Night bonuses are real evidence of real power. They’re also a reward signal that quietly teaches you to chase one outcome. In reinforcement learning terms, you’ve got a policy that gets a huge payoff from one action and learns to over-select it, because the big reward is loud and the slow accumulation of positional losses is quiet.
Bot builders do the same thing when we tune for the demo. The impressive path gets rewarded in every meeting. The boring failure modes — timeouts, ambiguous user intent, the request that needs three clarifying questions instead of one confident answer — never show up in the highlight reel. So we build agents that are spectacular on the happy path and get submitted in round two.
Signs your bot is bonus-hunting:
- It takes the aggressive action (write the file, call the API, commit the change) before it has confirmed state
- Its wins are dramatic and its losses are slow and quiet
- You can describe its best case in one sentence but not its failure handling
- Evaluation covers the scenarios it already does well
What a more strategic manner looks like in code
The fix isn’t to lower the power. Nobody tells a knockout artist to stop punching. The fix is gating.
Add a state check before the big move
Your agent’s most expensive, least reversible action should sit behind a precondition. Has it verified the file exists? Is the user’s intent unambiguous? Is this the right round for this punch? A cheap check before an irreversible action costs milliseconds and saves entire sessions.
Score the positions, not just the finishes
If your only metric is task completion, you’re measuring knockouts. Track the stuff that happens on the way there: how many turns to resolution, how often the bot had to backtrack, how many times it guessed instead of asking. Those are the decision losses, and they compound.
Build a losing-position policy
Fighters drill what to do off their back. Most agents have no equivalent. When your bot’s first approach fails, what happens? If the answer is “it tries the same thing with slightly different parameters,” you don’t have a recovery policy, you have a loop. Define the behavior explicitly: diagnose, switch approach, or escalate to the human.
Still on the roster
The part I actually respect is the framing. “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.”
That’s the right read on a bad stretch. Two losses didn’t delete the power that earned those bonuses. They exposed a gap between the capability and the decision-making around it, and a gap like that is fixable in a way that missing talent isn’t.
Same goes for the agent you’re debugging right now. If it can do the hard thing but keeps doing it at the wrong moment, you don’t need a new model. You need better conditions on the trigger.
October 10 in Las Vegas will show whether the adjustment took. Your next deploy will show whether yours did.
🕒 Published: