25 invalid decisions across 30 rounds. That’s the raw count from AI agents hammering a public goods game, where a multiplier juiced balances past an arbitrary cap—locking players out of their own success.
Look, incentive design stress-testing with AI agents isn’t some sci-fi gimmick. It’s already surfacing the kind of boundary bugs that tank real-world token launches (remember the DeFi exploits that drained $600 million in 2022? Underspecified rules, every time). This tool—Agent 006—lets anyone, zero programmers included, throw natural-language specs at Claude-powered pipelines and watch adversarial bots exploit the gaps.
How One Ambiguity Broke the Economy
The spec seemed solid. Five agents, 100 starting tokens each, 30 rounds of contributions to a public fund multiplied by 1.5, split equally. Collapse if gifts drop below 5 tokens for three straight rounds.
But it didn’t specify a max contribution beyond “0 to their current balance.” Claude’s extractor? First run, it slapped on a 100-token hard cap—matching starters, ignoring growth. Balances ballooned past that thanks to compounding returns. Boom: agents tried big pours, got rejected. 25 fails. Reporter screamed “parameter flaw: cap doesn’t scale.”
Weeks later, same spec. Different run. Cap jumps to 1,000 with dynamic clamps. Zero invalids. Clean sail.
There are 5 agents. Each round, each agent decides how much of their private balance to contribute to a public fund—anywhere from 0 to their current balance. The fund is multiplied by 1.5 and distributed equally. Agents start with 100 tokens. The game runs for 30 rounds. If total contributions drop below 5 tokens for 3 consecutive rounds, the system collapses.
That’s the magic—and the mess. Non-determinism isn’t a bug; it’s the stress-test’s secret sauce, forcing multiple interpretations of your sloppy English. Run it thrice, get three worlds. Cheaper than prod disasters.
I wrote that spec. Missed the gap. Tool nailed it.
Can AI Agents Outsmart Human Game Theorists?
Short answer: Not yet. But they’re damn good at boundary pokes.
Pipeline’s dead simple. CLI zap: npx tsx src/cli.ts –spec my-scenario.txt. Four Claude calls chain up: extract specs into JSON skeletons (flags ambiguities—fix ‘em), spawn JS sim engines, mint adversarial archetypes (greedy defectors, paranoid hoarders, whatever fits your rules), code their decision trees. Then simulate N rounds, invariant checks, post-mortem report.
Limits? Simultaneous moves only. Single actions per tick. No fancy sequencing. And yeah, Claude might hallucinate your economy wrong—it’s interpretation, not verification. But for protoyping DAOs, bonus pools, resource shares? Gold.
Take the ultimatum game. Proposer splits a pot; responder accepts or both get zilch. Classic Nash equilibrium crusher—fair splits win, greed loses.
Tool choked here. Generated proposer code looped infinitely on edge cases (pot=0?). Sim hung. Reporter spat errors. Fix? Tweak spec, rerun. That’s the loop: break, debug, iterate. Humans do it mentally; AI externalizes the grind.
My take? This beats whiteboarding. Forces you to spec tighter upfront. And in a world where crypto incentives bleed $10B yearly to exploits (Chainalysis 2023), early signals like these save stacks.
Here’s the thing—corporate hype calls AI “game theory 2.0.” Nah. It’s a flashlight in the spec fog, not Nash’s crystal ball. But wield it right, and you’re dodging the pitfalls that felled Terra Luna’s algo-stablecoin (vague peg mechanics, anyone?). Bold call: by 2025, every token drop will run variants of this pre-launch, slashing 30% of design-side hacks.
Why Non-Determinism Beats Deterministic Dullness
Same input, wild outputs. First public goods: flaw parade. Second: smooth. That’s not unreliability—it’s revelation.
Adversaries adapt per run. One batch goes full defector, starves the fund early. Another cooperates then flips. Wealth curves diverge: run #1 peaks at 450 tokens/agent before cap chaos; #3 hits 1,200 unchallenged.
Data dump from repo runs shows variance: average collapse round shifts 8-22 across five seeds. Invariants like “no negative balances” hold, but participation craters variably.
Skeptical? Fair. No theorems proven. But market dynamics scream value—web3 builders burn $millions iterating live. This? Pennies per run, issues surfaced Week 0.
And the ultimatum probe? Post-bugfix, proposers offered 40-60% splits consistently. Responders rejected below 30%. Equilibrium held, but one archetype—a “punisher” bot—nuked three rounds by spite-rejecting 49/51 offers. Human insight: add reputation layers next spec.
Single caveat. It’s Claude-locked today. Swap in GPT? Recode extractor. Open-source repo begs for it.
This isn’t replacing economists. It’s arming them with sim fodder no intern could crank fast enough.
The Bigger Market Play
AI agent swarms for design validation? Early innings. But zoom out: incentive markets underpin LLMs (RLHF rewards), blockchains (staking yields), even ad auctions. Flub the rules, watch value evaporate.
Parallel: 2008 quant funds stress-tested CDOs with Monte Carlos—saved billions. AI does that for tokenomics now, at hobbyist scale. PR spin says “autonomous economies.” Reality: exploratory wrench, not oracle.
Run your own. Fork the GitHub. Spec a referral bonus gone wrong. Watch.
Worth the hype? For indie builders, yes. VCs? Mandate it in pitch decks. Enterprises? Too toy-like—needs async moves, multi-resource states.
But don’t sleep. Tools like this scale with model IQ. Claude 3.5? Tighter sims. o1-preview? Smarter adversaries.
🧬 Related Insights
- Read more: Agentic AI’s 2026 Scale-Up Nightmares: 5 Roadblocks Killing Prototypes
- Read more: Google’s Traumatized AIs: Meltdowns in the Machine That Could Derail Your Workflow
Frequently Asked Questions
What is Agent 006 AI tool?
Agent 006 is an open-source CLI that turns plain-English economic specs into JS simulations run by adversarial AI agents, flagging incentive flaws early.
How do you stress-test incentives with AI agents?
Write a text spec, run npx tsx src/cli.ts –spec yourfile.txt—AI extracts, simulates, and reports failures like invalid actions or collapses.
Does AI incentive testing replace game theory?
No, it’s a fast exploratory check for ambiguities, not formal proofs—run multiples for best signals.
