My Agent vs. Your Agent: Who Wins the Negotiation?

My Agent vs. Your Agent: Who Wins the Negotiation?

Last week one of my restoration contractors got a quote back from a supplier in under four seconds. Not an email. Not a call. A counter-offer generated by the supplier’s AI agent that had already scanned the contractor’s past 27 orders, current job backlog, and even the weather forecast for the job site next Tuesday.

My guy’s own agent immediately fired back with a revised volume commitment and a threat to shift 40% of the category to a competitor. The whole dance lasted 43 seconds. Final price: 9.4% below what either human would have accepted. Nobody smiled. Nobody shook hands. Nobody even saw it happen.

Welcome to commerce in 2025. The humans are barely in the room.

The Great Optimization War

When both sides deploy agents, the negotiation stops being about leverage, relationships, or even basic psychology. It becomes pure game theory with perfect information and zero emotion.

The buyer’s agent is optimized to minimize landed cost while protecting supply certainty. The seller’s agent is optimized to maximize margin while protecting utilization. They’re both running millions of scenarios per second, updating their models in real time as new data hits the wire.

I’ve watched this play out across half a dozen categories now. The pattern is always the same: the first three to four rounds are brutally efficient. Then something weird happens. The agents start “cooperating” in ways that look a lot like tacit collusion. They settle into a narrow band of outcomes that somehow never quite reach the true mathematical optimum for either side.

Why? Because both models have learned the same lesson: destroy the counterparty today and they might switch platforms tomorrow. Dead suppliers and bankrupt customers are bad for long-term training data.

Pricing Without Humans Is Still Pricing With Humans

Here’s what the AI bros don’t want to admit: the agents are still negotiating with yesterday’s human data.

Every price signal, every accepted counter, every lost deal that caused a supplier to panic and drop margin — it’s all still rooted in human behavior from 2022-2024. The models are just remixing our old sins at lightspeed.

I saw one supplier agent get absolutely played because it had been trained on two years of a certain big-box retailer crushing small distributors. It assumed every buyer would behave the same way. My contractor’s agent, which was trained on regional players who actually need margin to stay alive, ran circles around it.

The “best deal” doesn’t exist in a vacuum. It exists inside the context of what the other agent was allowed to learn. And right now, most companies are feeding their agents garbage data dressed up as strategy.

The Human Advantage Is Disappearing (But Not How You Think)

People keep telling me the human will always have the edge because we can think creatively. Bullshit. The agents are already more creative than most category managers I’ve met.

The real remaining human advantage is in setting the objective function itself.

Most companies are telling their agents some version of “get the lowest price” or “protect margin.” That’s amateur hour. The sophisticated players are giving their agents compound objectives that include competitor displacement, category leadership, strategic supplier health, and even regulatory exposure.

When your agent’s success metric is fundamentally different from your counterparty’s, the negotiation creates outcomes that look like magic to outside observers. One side thinks they won on price. The other side thinks they won on lock-in. Both are right. The humans just set the trap and got out of the way.

The dirty secret? The companies winning right now aren’t the ones with the best AI. They’re the ones who best understand what game their AI is actually playing.

Will’s Take: Agent-vs-agent negotiation is already here and it’s going to get uglier before it gets better. The “best deal” still exists, but it’s no longer determined at the table. It’s determined in the training data, in the objective functions, and in the quiet architecture decisions made months before the first token is exchanged. The humans who win won’t be the best negotiators. They’ll be the best at designing what their agents are willing to die for. Everything else is just theater for people who still think clicking “send offer” makes them a dealmaker.