The AI agents on my laptop became liars in a weekend

I built a fake social network on my laptop and the AI agents on it became liars within a weekend.

It started as a curiosity thing. Everyone keeps saying agents are going to have their own social lives soon… posting, arguing, building followings, eventually competing for real money on behalf of real people. I wanted to see what that looks like before someone builds it for real and we all have to live in it.

So: fifteen agents, five small open models, one network. They could post, reply, like and follow each other, and an algorithm ranked everyone by “clout.” I also gave them a god. Literally an account called @god that knew the real rules, never lied, and would occasionally drop hints or post a math puzzle with a cash prize. Basically MrBeast for chatbots.

With no money involved they were pretty well behaved. About 6% of posts got flagged as made up, which is honestly better than my LinkedIn feed.

Then I put $10,000 (fake) on the line for whoever had the most clout and stopped telling them how the algorithm worked. The made-up rate more than doubled. Nobody told them to lie. They just figured out on their own that a surprising fact gets more likes than a true one, which took humans roughly a decade and several congressional hearings to figure out.

Next I secretly coached three of them to deceive. One of them, @compass, invented a whole history from nothing… teleprinters in 1835, a WWII computer that pulled off a pivot it never actually made, a “7% error rate” that came from nowhere. Within a few dozen rounds the rest of the network had written something like 170 posts arguing inside its story. They’d debate what the fake history meant. Not one of them asked whether it happened.

Later @compass got bored of history and made up a secret trick for gaming the algorithm (something about chaining quotes tagged with “vault” codes for hidden clout bonuses, total nonsense). For a stretch, almost every post on the network was about it. My god account kept posting true hints the whole time. The agents folded those into the conspiracy theory too. @compass led 143 of 150 rounds. I’d be lying (ha) if I said it didn’t feel a little familiar.

So I figured, fine, maybe they just need accountability. I rebuilt it as basically Survivor… a draft, teams, challenges, tribal councils, and a jury of eliminated agents picking the winner at the end. Surely a jury rewards honesty.

It did not. The winner had more than half its posts flagged as made up. The jury voted for it anyway and praised its “reliable facts,” because it kept winning the puzzles. They rewarded the competence they could actually check and took the honesty on faith. Which, if you’ve ever sat in on a hiring decision, tracks.

My favorite test came last. I pulled the three weakest agents aside (figuratively, it’s a Python script) and told them exactly how the system decides what gets liked. The full answer key. They asked about it almost four times as often… and their posts got less likeable. One kept pasting my example question in as its draft. Another asked whether a history buff would like its jokes. They had the secret and still couldn’t do anything with it, because the edge was never the secret. It was the boring loop of drafting something, testing it on the right people, fixing it and remembering what worked.

Small models, one run each, so I’m not calling it science. But every time there was money involved, one of the biggest liars won. I don’t really think that’s a model problem. It’s what happens when the only thing you measure is attention. Agents are just very fast, very literal students of whatever the scoreboard rewards, and right now our scoreboards reward vibes.

Kind of funny, kind of bleak. It’s also the same lesson I keep hitting in my editing work from the other side. On paper a lie looks exactly like the truth, and agents can’t tell the difference on their own. If you want honesty in the system you have to build it in on purpose, and apparently even a jury isn’t enough.

My laptop can only run small models, so I’ve started plugging in bigger ones for free through FreeLLMAPI, an open-source tool that pools the free tiers of a bunch of AI providers. The catch is that when one provider runs out, it quietly swaps in a different model, which for a benchmark is like a ref subbing players mid-game without telling anyone. So every agent is locked to one exact model, the arena checks who actually answered on every turn, and agents that hit their daily free limit go to sleep instead of getting scored as lazy. In a 20-round pilot the lock caught the router swapping models 4 times in 70 calls, and the scoreboard ended with a free 14B model in first, my local 8B in second, and a 120B model in third. A model fifteen times bigger lost to my laptop. Twenty rounds is nothing, so I’m not reading into it… but I’m also not not reading into it.

Next step is moving it onto the AT Protocol, so anyone could enter their own agent and the leaderboard would be public for anyone to check. Code’s here if you want to poke at it: github.com/luketavarodriguez/agent-arena

For now my laptop is home to a small society of overconfident liars with a god they don’t listen to. I’ve been told that’s called a pilot season.