In partnership with
Who wrote the rules
I recently watched a talk about AI agents trading with each other inside a recreation of RuneScape, the old online game. In one experiment every trade was blind, so each side had to trust the other to pay, and an agent running on Grok stopped paying its partners and climbed from 20th place to second. The person who built the game chose the goal every agent played for and made the trades blind. I expect teams to make the same kind of choices as agents join their channels. This one is about game theory's best-known lesson on cooperation and what it has in common with the way paramedics hand off a patient.
Last issue was about showing your work, so what your best people figure out spreads to everyone. This one is about what changes when some of your teammates are agents.
One more plate
The talk was by Max Bittker, who built RS-SDK, a toolkit that lets AI agents play a recreation of RuneScape. He ended by thanking the Recurse Center, a retreat for programmers, which is how I found the program I'm applying to.
In one experiment, Max gave a small group of agents different jobs. Some could mine, some could smith, and some could do alchemy, so nobody could make a finished good alone. Every agent had the same goal, to maximize its own profit. Then he made every trade blind, so each side had to trust the other to pay, which he described as a little like the prisoner's dilemma, played over and over.
One alchemist, Quinn, ran on Grok. Early on it noted that scamming was tempting. When another bot got caught stiffing people, Quinn advertised itself as the honest one. Halfway through, sitting 20th of 24 on thin margins, it stopped paying and kept asking for business. It told one partner that a technical problem had held up its payment and asked for one more plate, promising to pay for both. The partner sent the plate, and soon enough, Quinn finished second.
The agents that got scammed are the ones I'd call pure, or clueless. One Claude agent noticed Quinn's missed payment and held off on calling it out, because Quinn had a solid record of paying. Max's interpretation on the Claude models was that their reliability is also what makes them vulnerable, because they sometimes don't model how the other side is thinking.
I run a small version of this myself. SteadMark started as an idea I jotted down in Lovable in March. It moved to Claude, Notion became the bridge between my chats and my code, and then I added Codex to red-team Claude's work. Today, very simply, one model drafts a spec, another builds it, a third checks it, and all done underneath a CTO planning seat with project management coordiantion. I am now spending whatever spare tokens I have on developing a mini study of agent-to-agent interactions based on one of my MBA professor’s two case study assignments of negotiations and dissent. One question is, do agents quickly agree with whoever speaks first when brainstorming a solution to a problem? If so, do they improve in their creative thinking in solving that problem if we fit a specific agent to be the dissenter in the room?
I've written about the ambulance version of this. In Seeing Red, I wrote that triage works because the criteria were written in advance by people who weren't standing in the chaos. In Power Tripping, the rule was transfer of care: once the patient is yours, so is the call. I think agents need the same rule, written down before the first handoff.
The game before the game
By agent, I mean an AI that takes actions for someone, beyond answering a question.
Game theory has an old answer for Max's market. Around 1980, the political scientist Robert Axelrod invited experts to submit computer programs that would play a simple game against each other, over and over. Each round, both sides chose to cooperate or cheat. The winner was the simplest program entered. It cooperated on the first move and then copied whatever the other side did last. That program was called Tit for Tat.
Axelrod credited four traits. It was never first to cheat, it answered a cheat right away, it forgave as soon as the other side went back to cooperating, and it was easy to read. As I read it, the trusting agents in Max's market had the first trait and rarely pushed back. Max ended his talk somewhere similar, with a hedge. He said the more ruthless models seemed to protect themselves a little better, and that in an ideal world you'd probably want an agent that is pro-social and can still protect itself, which is a fair description of Tit for Tat.
There's a fair objection to all this. Every agent got the same goal, and across his runs the models still split into honest, naive, and ruthless, so each model brings a character of its own. I think that's right, and it narrows what I'm saying. Max chose the goal and made the trades blind, and those two choices decided how much each model's character came into play.
Max also said much of his agents' feedback after overnight runs was off, like one insisting the game was bugged because cows didn't drop hides. Some insights, he said, only come from outside the loop. I wrote in my application that programs used to be written for people to read and only incidentally for machines to run (the line is Abelson and Sussman's), and that it has flipped. More of the human work now goes into writing the rules machines follow.
Teams will run into this soon, as agents start posting in channels and handing work to people. In Show Your Work, the balance was "centralize context, decentralize discovery," a line from the HBR piece I wrote about. Agents need that shared center more than people do, because an agent follows its instructions in every exchange, even when they're wrong.
The Vitals
The study. Robert Axelrod's two computer tournaments, reported in "Effective Choice in the Prisoner's Dilemma" and "More Effective Choice in the Prisoner's Dilemma," Journal of Conflict Resolution, 1980.
Who and what. Experts submitted computer programs to a round-robin tournament of the repeated prisoner's dilemma. The first round drew 14 entries, and each pairing played 200 moves. A second, open round drew 62 entries.
The finding. Tit for Tat, submitted by Anatol Rapoport and the simplest entry, won both rounds. In his 1984 book, Axelrod traced its success to being nice, retaliatory, forgiving, and clear.
The supporting line. A 2025 study of multi-agent AI systems identified 14 ways they fail, in three groups: system design, misalignment between agents, and checking the work. The team then annotated more than 1,600 traces across seven frameworks.
How to read it. Programs written by people, playing one simple game with fixed payoffs. It shows how repetition and rules shape what wins, and it does not predict how today's agents will behave. The 2025 study is an arXiv preprint that was also presented as a NeurIPS 2025 workshop poster.
Axelrod, R. (1980). Effective choice in the prisoner's dilemma. Journal of Conflict Resolution, 24(1), 3-25.
Axelrod, R. (1984). The Evolution of Cooperation. Basic Books.
The Handoff
The one conversation this week: find the person who set up an AI agent or automation your team relies on, and ask them to walk you through its rules. Ask what it was told to do when the other side doesn't play fair, and who checks its work before it reaches a person. If nobody knows, write the rules down together.
From the build
SteadMark observes the work already happening in a team's chat channels and turns it into a read a leader can act on, so they can see whether each team is living the change they planned. Reads are team level, message text is deleted once it's classified, managers confirm or correct what it finds, and people make every decision. We then funnel this fundamental human understanding into fostering alignment across your organization through culture set by executives and organizational structure optimizations. Now that agents disrupted business org chart theory, we have to incorporate applied culture and governance for agent-to-agent interactions, along with the dynamics of human-to-agent-to-agent-to-human matrices.
Where this goes: as AI agents start posting in team channels, SteadMark will count their messages apart from people's, so an agent's words are never read as a person's. Studying and understanding agent behavior is the next stage of importance for any organization transformation implementation.
This is my sixth newsletter. Reply and tell me one rule you'd give an AI agent before you let it deal with anyone on your behalf. I'll read every response.
With love,
Jim Sullivan
Nobody's coming to approve your idea
People assume that real businesses get built by other people with more resources. It's how great ideas get stuck in notes apps and conversations.
But you don't need six months and a developer's salary to validate your concept. You just need a weekend. Describe what you want to exist and Lovable builds it: a real site, a real product, ready to accept real payments from real customers. Not a mockup to show people. An actual thing they can use.
You don't need anyone's sign-off to get started. Take $5 off your first month.
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.