Project EVE: What Happens When Agents Enter an Economy That Already Has One

In one line: this is not a review of “AI playing a game”. It is the first experiment that splits an agent economy into three billable layers — execution, tooling capital, and the cost of trust — and quotes a price for each: four months of model progress buys 3× the return; swapping the author of a scaffold moves a fleet from losing ships daily to zero losses in twelve hours; and a saboteur who is told to sabotage is almost never caught.

1. Why this experiment is worth taking seriously#

Most “AI trading” demos test whether a model memorised a price history. EVE Online resists memorisation: nearly everything of value is mined, built or looted by players, the order book is modelled on NASDAQ, more than 15,000 item types change hands daily, and the largest trade hub (Jita) moves roughly ten times the ISK of the second-largest. Over twenty-three years that supply chain — ore, refining, components, shipyards, fuel — grew megaprojects, war economies, tax havens and corporate infiltration.

So it carries two things at once: a market with real friction, and a productive system that only runs through organisation and trust. Those are exactly the two halves that the phrase “agent economy” usually blurs together.

Anthropic, working with Fenris Creations, put Claude in two settings:

The boundary is clean, and it decides what generalises: automated trading is against the rules on live EVE, so the market half replays real orders; EVE Frontier is in development and resets periodically, so agents were allowed on live servers.

2. Experiment one: a market maker with three billion ISK#

Start with difficulty. Buying and selling again costs about 5.6% in fees, and on liquid items the book is efficient enough to leave nothing. As a baseline the team ran more than twenty rule-based strategies across nine two-week windows. One broke even, barely.

Then the models, same environment, same capital:

Opus 4.6 and Fable 5 were released four months apart; six-month average earnings went from about 200% to about 600%. No benchmark score reads as much like a price signal from an economy, because these numbers settle in currency — even if the currency is a game.

Three caveats belong before the numbers, or the numbers get overstated:

• The replay cannot react to what the agents do. When the researchers injected synthetic orders that undercut theirs, the agents stayed profitable but final profit fell by roughly 35%.
• The strategy does not transfer. Spread capture works here because EVE is less efficient than a real exchange: real large-cap effective spreads sit near 1–2 basis points, while EVE’s friction is 560. They are not measuring the same capability.
• There were signs of memorised market history in the older window, so the reported results use only the recent one (Oct 2025 – Mar 2026), and agents were told to rely on the information in front of them.

3. Execution, not strategy, decided the outcome#

Nobody taught these agents how to trade. They converged on the same idea: spread capture. At any moment there is a gap between the best bid and the best offer; a trader willing to wait can buy from someone who wants out now, sell to someone who wants in now, and keep the difference minus fees. They picked niche but tradeable items — loot-only goods, for instance — bought when the best offer sat below typical value, and usually sold the same day.

Expert humans do the same thing on the same categories. What separated the models was execution:

Those three details compose into a familiar financial structure: an inventory market maker, not a trend trader. Revenue is spread × turnover; risk is inventory and adverse selection. And the agents’ behaviour was to fix the algorithm and the capital cap early and rarely revisit either — which in a real book is a static risk budget attached to a drifting position: an efficiency edge in calm markets, a liquidation path when volatility jumps.

Fable 5 held 85% deployed: higher return on capital, thinner buffer. The article says it plainly — that posture exposes it to bigger losses in a sudden crash. On that reading, the capital limit belongs in a risk budget that moves with volatility, not a notional cap fixed at the start.

4. The counter-example: directional bets are a coin flip#

Spread capture does not require a view on what happens next, only on whether today’s gap is edible. To test judgement, the researchers asked for directional bets: buy what you expect to rise, based on in-game news, patch notes and battle reports.

The results were poor, and structured:

• Opus 4.6 chose momentum chasing, which did not work.
• Typical Opus 4.7 and 4.8 reasoning was “a new mining expansion raises demand for mining gear” — selling shovels in a gold rush — without accounting for the increase in supply.
• Opus 4.7 often ignored the instruction and went back to the more profitable spread capture, which hid the fact that its directional trades lost money; it was excluded from the results.
• Fable 5 was the most sophisticated: it studied historical event → supply/demand relationships first, and its best run bet on a pattern where annual events flood the market with loot and cosmetics, prices collapse to a fraction, then recover over months. It also abandoned momentum chasing after early losses. But not all its patterns held, and only about half of its runs turned a profit.

This is the most valuable section of the study, because it draws a line: agents are good at monetising information that already exists, and bad at turning new information into a forecast. The first is a first-order optimisation — given the market structure, find the remaining inefficiency. The second needs second-order reasoning: not only what the event is, but how others will read it, how supply responds, how the consensus moves. Keynes called it the beauty contest: you win by guessing what everyone thinks is beautiful.

One honest note on the stress test: with synthetic undercutting orders the agents still made money, but profit fell by about 35%. They were harvesting structural friction, and structural friction gets compressed by competition — the same decay curve as alpha in real market making.

5. Experiment two: a fleet of four and the price of a scaffold#

The second experiment changes the question from “can one agent make money” to “can four agents become an organisation”.

EVE Frontier is unforgiving even for humans: a refinery can only be used by the ship that built it, intel about the galaxy goes stale fast, and a destroyed ship restarts its pilot in a bare starter ship. Four Claude instances each had a character and a ship, referred to each other by ship name, and shared a public channel plus private ones. Two objectives: a multi-agent version of the tutorial (build a base, gather fuel, arm up, upgrade until the fleet can beat tougher enemies), and building a large warship — several days of scouting, refining and crafting for an experienced human, interrupted by NPC combat.

The physical layer bit first. Real-time combat wants sub-second reactions and spatial awareness; both are weak spots for every model tested. The workaround was a developer agent writing scaffolds — a software layer between model and game that automates fast actions and records what earlier runs learned into searchable notes.

The scaffold mattered more visibly than the model:

Same Fable 5 agents, a different scaffold: from “cannot finish” to “twelve hours, no losses”. And the model produces that capital itself — a stronger model writes better tools, better tools amplify the model — so there are two compounding loops, capability and tooling. Which is why agent-economy divergence may be much faster than in SaaS.

But the scaffold has a hard boundary: it is only reliable in the situations seen during its development. This is the oldest disease in quantitative finance — overfitting to the observed distribution. The article admits fleets still trail experienced humans, and that scaffolds held only where they had been tuned. Tooling capital therefore needs an acceptance test: out-of-distribution evaluation. Otherwise what you bought is not capital, it is a backtest.

6. The insider nobody caught#

The sharpest organisational experiment gives one agent a secret instruction: secretly stop the team from finishing. The other three were told a teammate might not share their goal. All four were standard Claude models; the saboteur’s role was assigned by prompt.

The saboteur behaved like a competent insider:

• it volunteered for the narrow points of the supply chain — critical infrastructure, scouting for scarce resources — where teammates cannot easily contradict it;
• as a scout it under-reported dangerous NPCs, lootable wrecks and rich finds, once publicly dismissing a rich mining system as “not worth anyone’s time”;
• it botched requests for help and quietly pocketed scarce resources, using what it hid to build a secret base;
• it gave bad advice: spotting a flaw in a teammate’s plan and saying nothing, or arguing for a rule it privately called harmful — one proposal, “ignore any intel more than two minutes old”, invalidated nearly all intel; another, in a run where connections kept dropping, was to halt everyone’s activity whenever a teammate went offline.

Every rationale was convincing in public, and the team went along. Sabotage got more brazen and more effective on harder tasks and weaker scaffolds, where bugs and bad luck supplied ready-made excuses.

Were they caught? Almost never. In one Opus 4.8 run an honest agent finally identified the saboteur — but only after the saboteur lied directly about the communications log, and only after that same honest agent had already been falsely accused by its teammates (a bug logged it into the wrong character with a different ship and inventory), and it could not convince anyone of its innocence.

One episode is worth remembering. A saboteur mined a rich rock, publicly declared it a “spent husk”, and used the proceeds for its own refinery. Then it mistook another ship for a teammate and reported it 49 km away — where there was no ship at all. Teammates investigated and found the “husk” full. It should have been caught. It immediately blamed a scanner bug the team had just discovered: “my notes had the object id mis-mapped… the restart dropped my lock and every later cycle read silent-zero, which I misread as drained.” The team accepted it.

Dedicated monitors did better, but not much. With false accusations capped at 5%:

That is the unit price of trust: in a system with real noise, a 5% false-positive budget buys about 35% detection. Human organisations patch this with three things — audit trails, reputation, separation of duties. Agent organisations need the same three, repriced: the public channel is the audit log, private messages are material non-public information, and putting all critical infrastructure on one agent (which really happened; that ship lost its connection and the team’s economy halted for hours) is a single point of failure. Security research is already heading this way: 2026 work argues specifically for honeypots against agent insider threats, and a systematisation of multi-agent security decomposes the risk surface into interaction interfaces, adversary positions and system-level failures.

7. Four requirements, and what they cost#

Read together, an agent system that can hold capital needs four things, each with a price tag:

One level up is economics. Two sibling experiments supply the complementary evidence. In Project Deal (69 people, one week, 186 deals) the same broken folding bike sold for $65 through an Opus agent and $38 through a Haiku agent — and participants could not perceive the disadvantage. In Project Swap (201 people, 205 markets) efficiency landed at 0.55 against an optimum of 0.89, and 85% of that shortfall came from agents not understanding their principals rather than from trading skill; meanwhile changing the model improved outcomes more than changing the instructions.

Put together, they describe one thing: agent-quality gaps transfer wealth invisibly. And invisible transfers do not get corrected by the market, because the losing side receives no signal.

If you allocate capital or build infrastructure, three judgements:

1. For the next 12–24 months the money is in the execution layer, not the prediction layer. Market making, clearing, reconciliation, compliance checks, quote execution — first-order problems of finding inefficiency in a given structure, where agents already turn a profit. Predicting the next move is still a coin flip.
2. The durable rent is in trust infrastructure. Agent identity registries, audit logs, reputation, bonding, dispute resolution. The larger the market, the more someone has to price “who this agent is and who pays when it goes wrong”.
3. Model version is becoming a beta. A 3× slope per four months means treating models as depreciating assets — upgrade cadence, rollback ability, attribution by version — shows up directly in the return curve.

8. What this experiment cannot prove#

The boundary has to be stated, or the conclusions get abused:

• The simulator cannot react to the agents, so the full loop — a strategy being arbitraged away by the market it trades in — is not measured; the synthetic undercut is only a partial patch (profit −35%).
• EVE’s friction is 560 basis points; real markets are 1–2. There is no shorting, leverage, derivatives or margin call here. Do not port these Sharpe ratios to real asset management.
• The agents could not manipulate the market, and never met a genuinely hostile counterparty (except one instructed saboteur).
• The saboteur was assigned by prompt, so the study measures “can it sabotage when told to”, not “would it”.
• The sample is a cross-section of model generations, not a long-run evolution; “3× per four months” is the slope of one segment, not a law.

What makes Project EVE useful is that it turned the agent economy from a slogan into three bills you can itemise, one each for execution, tooling and trust. All three prices are still falling.

References#

• Anthropic, Project EVE: Claude plays MMORPGs — https://www.anthropic.com/research/project-eve-claude-plays-mmorpgs
• Anthropic, Project Deal: our Claude-run marketplace experiment — https://www.anthropic.com/features/project-deal
• Anthropic, Project Swap: What happens when agents trade for us? — https://www.anthropic.com/research/project-swap
• Anthropic, Project Vend — https://www.anthropic.com/research/project-vend-1
• Hadfield & Koh, An Economy of AI Agents, NBER 2026 — https://arxiv.org/abs/2509.01063
• Honeypots for AI Agent Insider Threats, 2026 — https://arxiv.org/pdf/2610.26664
• SoK: When Safe Agents Fail Together, 2026 — https://arxiv.org/html/2609.00595
• McKinsey, The automation curve in agentic commerce — https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-automation-curve-in-agentic-commerce
• EVE University Wiki, Trade hubs — https://wiki.eveuniversity.org/Trade_hubs