AERead (AgentEcon Readiness) is a benchmark and open environment that asks whether an LLM can actually operate in an economy: how much of the value available to a group of trading agents gets realized when it takes a seat, rather than whether it can finish an isolated task.

What is the Aggregate Efficiency Ratio (AER)?

It is the score defined beside the leaderboard: the share of the best possible outcome the market actually reaches. Two things it is not. It is not the tested agent’s own profit; a model can bargain hard, take a large share, and still score low if the market as a whole leaves value unrealized. And it is not a percentage of some ideal economy: the best possible outcome is recomputed for each case from that case’s private values, which is what makes scores comparable across very different markets.

What is the frontier problem AERead identifies?

Two results side by side. In the live five-seat market, agents that can see what their counterparties value make real progress, and still leave most of the achievable value on the table; that gap is measured against a computed best case, not asserted. (Saturation shows up only in the isolated scripted probes, where nothing is hidden and no one reacts.) Once information hides (finding partners, pricing private goods, getting consent), realized value falls back toward a trivial rule-based baseline, and deals never grow past two parties. The same frontier covers what strong institutions normally absorb for people: getting many parties to agree at once, and trusting a deal will settle.

How does the evaluation market work?

Each case is a five-seat market. The model under test takes one seat; the other four are a fixed LLM panel, pinned to the same settings for every run so each candidate faces an identical opponent. Agents negotiate in free-form language; their offers and agreements are parsed, checked, and settled on a shared ledger. Whatever the market ends up with is then scored as AER, defined above.

Do institutions matter as much as models?

It looks that way, and the pre-registered ablations already put a number on it (one model, one panel, development scenarios): changing a single rule of the market (whether funds are held in escrow, how agreement is obtained, how much agents can see about each other) moved the score by .05–.13 AER, comparable to the largest gaps between models on the held-out board. That is one implication among the findings, not the mission: the core stays the benchmark and the open environment, and a track that tests this head-on is planned.

Is AERead about ranking models?

No. The load-bearing result is the shape of the gap: realized value stays far below the best possible outcome, and the gap widens as information hides. The same pattern appeared in every model tested in pilot V0. The case suite is deliberately selected, too: cases pass an admission screen run independently of any model provider, which keeps settings where decentralized agent intelligence is the binding constraint and rejects already-solved market machinery (posted price, order book, matching). So pooled scores are stress-test statistics for decentralized exchange, not an estimate of how efficient agentic commerce is overall.

How is AERead different from task-completion benchmarks?

A task benchmark scores whether the model finished the job. AERead scores how much value a group of trading agents actually captures while the candidate is one of them, under realistic information conditions, so it measures economic competence, not task completion.

What do these cases correspond to in the real economy?

Each case family is a familiar economic activity. Bilateral exchange is the simplest deal, a used-car sale between neighbors with terms in plain view. Multiparty clearing is a supply chain or a housing chain, value that only exists if several trades close together. Discovery is hiring, sourcing, or selling a house, where most of the work is finding the counterparty at all. Consent is a salary negotiation or an acquisition, pricing what the other side will not show you and getting a real sign-off. Deferred settlement is trade credit, pay first and trust the counter-leg arrives. The cases emphasize these frictions rather than isolate them, and the steps where models break (finding partners, pricing hidden values, extending trust) are the ones that carry most of a real economy.

What does agent-to-agent (A2A) mean in AERead?

The counterparties are live LLM agents that negotiate back, not a scripted environment. They are a frozen panel, pinned as described above, so what varies between runs is the candidate and not the opposition. The consequence is the point: a good plan earns nothing unless the other side agrees to it, and the other side has its own reasons not to.

Is any model close to solving the hidden-information frontier?

No. In pilot V0, once information is hidden, every model tested falls back toward a trivial rule-based baseline. In the live five-seat market (not the isolated scripted probes), the strongest agent in one seat gets the market to about 0.38 of the best possible outcome when values are visible, and about 0.16 once partner discovery is hidden.

AERead is developed by Chenyu Li, Zeyu Sun, Yuchen Fang, Yuhan Cao, and Guangting Yu.

AERead: Can an LLM act in an economy?

00 / The 30-second version

How does the test work?

Here is one run of the market, in about thirty seconds.

01Five seats, private valuesFive traders take the five seats in one small market, and each one knows privately what the things it holds, and the things it wants, are worth to it.
02Information is hiddenNobody sees the whole board: depending on the situation, an agent may not know what the others value, or even who is out there to trade with.
03They talk it outThe traders make offers, haggle, and accept or refuse in ordinary English.
04Check and scoreTalk is not a trade: an agreement only counts if the goods exist, the budgets hold, and everyone whose things move said yes, so a confident talker cannot score points for imaginary deals. The deals that pass go through, and the value created is compared with the most those five traders could possibly have created.

What a score means: imagine a market where the best possible set of trades would create $100 of value. If the deals the agents actually strike create $38 of it, the score is .38, meaning about 38% of the value on the table was realized. That share is the project's one number, the Aggregate Efficiency Ratio, or AER. The two bars above show one model, the strongest one tested in this pilot. Both scores come from the held-out situations: private test versions that are never published, so no model can practice on them. With everyone's values in plain view, the market reached .38. When the model first had to find its own trading partners, .16. The rest of this page is about why the second bar is so much shorter than the first.

00 · b / The environment

A general environment for multi-agent economics.

The five cases are one configuration of something more general. Worlds are generated from a seed, and 26 fields set the rules of the economy: who can see whom, who may contact whom, how consent is obtained, whether a swap must clear all at once. Change one field and you have the same world under two regimes, which is how every claim on this page gets tested.

The piece that makes any of it possible is the compiler: agents write ordinary sentences, and the environment has to turn those into transfers it can actually execute. The compiler and the verifier are seats in the environment, not part of the scorer, so a deal that cannot be grounded is a deal that did not happen.

Text the compiler cannot ground, or a row the verifier rejects, settles nothing: no welfare is created, and the ceiling it is measured against does not move. That is the deployment-realistic choice, and it is why a fluent agent whose deals never execute scores badly.

In short

What it measures. How much of a market's welfare is realized when a model trades against other agents: economic competence, not task completion.
The finding. Performance is lowest where hidden information binds: about .38 of the best possible outcome with values visible, about .16 once discovery is hidden. Hiding counterparties alone costs the tested model two-thirds of its score; the cases vary more than visibility, so full attribution stays open.
The shape. The same qualitative gap appeared in every model tested in pilot V0, across providers and price tiers.
How it's scored. Five-agent agent-to-agent markets; an oracle grades realized ÷ achievable welfare (the AER); results are checked on private held-out seeds.

01 / Individual decisions

Can one agent find the right move on its own?

One model, alone, against fixed rule-based counterparties that never adapt. A solver computes the best outcome the situation allows, and the agent is graded against it. This isolates plain deal-making skill from everything that live negotiation adds.

Why score the market, not the agent? Because in these situations the two come apart. Self-interest does not add up to the best outcome: holding out for a better price, not bothering to look for a partner, and refusing to move first on credit are each sensible for the agent doing them, and each leaves value uncreated. So a score built on an agent’s own haul would reward exactly the behaviour that shrinks the pie. It would also barely separate the field: in this pilot every model took a similar share of the deals it joined, while the value those deals created differed several-fold. What varies between models is how much they create, not how much they keep, so that is what the benchmark measures.

AER (Aggregate Efficiency Ratio) is the score: the share of the best possible outcome that the market actually reaches. That best outcome is computed, not guessed: a solver is shown everyone’s private values and finds the arrangement that leaves the group as a whole best off. 1.00 means the market got there; 0 means it did no better than the starting allocation; a negative value means the trades left the group worse off. It scores the whole five-seat market, not the tested agent’s private take. A second score, graded against the best achievable without seeing what others know, is in development.

1.00 AER

Frontier performance on visible construction reaches the oracle ceiling.

current difficulty · saturated

4 / 4

Solved: procurement bundling, private-value auction, cyclic clearing, and coupled B2B award + re-split.

frontier-model probe · n=30 · attribution on the evidence page

−0.23

Surplus efficiency on hidden-reservation pricing; the scripted ceiling is not saturated here.

market-ext · n=20

market-ext is an extra set of market cases (bundled procurement, an auction where each bidder knows only its own value, cyclic clearing, a coupled B2B award, and pricing against a hidden reserve) run in scripted mode: one model against fixed counterparties that never adapt. These probes are a separate track from the agent-to-agent leaderboard on the evidence page; don’t read them as leaderboard entries.

Key question

Is it still hard when everything is visible?

Values and constraints are on the table, and no counterparty reacts. All the agent has to do is find a trade that leaves the group better off, before we ask it to read intentions, bargain, or coordinate with anyone.

A2Scriptedcontrolled capability cell

Key finding

Assembling the deal, yes. Pricing a hidden number, no.

These are the scripted probes: one model, fixed counterparties that never adapt, every value on the table. There, frontier models saturate the cases that ask them to assemble a deal. The open problem begins where a counterparty’s walk-away price is private and has to be inferred, then priced. This is the benchmark’s floor, not its headline; the live-market scores in the next section are far lower.

strong optimizerweak strategist

02 / Emergent A2A market

How much of the available value do live agents actually capture?

Now the counterparties are live LLM agents that negotiate back, reveal or withhold information, and settle on a shared, verified ledger. They are the frozen panel, and the restraint is diagnostic: the panel responds but only the model in the tested seat takes the lead, so whatever the market creates can be traced to that one model (every seat taking the lead is the arena’s question, below). The score becomes how much of the best available outcome the whole market reaches once it can react and once information is hidden, which is exactly where agents fall short of what they manage in plain view.

01Under testthe model being scored takes one seat in a five-seat market.
02Frozen panelthe other four seats are a fixed model pinned at temperature 0: reactive counterparts that negotiate, not scripts.
03Verified ledgerfree-form offers and consent are compiled, verified, and scored: the same arena pipeline.

A2A swaps A2Scripted's deterministic counterparts for negotiating LLM agents. The AER then measures how much of the available market welfare is realized when the under-test model holds one seat and counterparts can bargain and hide information.

02 · cases / What these cases are in the real economy

AI agents are starting to buy, sell, hire, and negotiate on our behalf. The five cases are not abstract puzzles; they are the five things that make a market a market, each a friction economics has studied for a century. Each case emphasizes one of them rather than isolating it.

case01 · bilateralMake a dealA used-car sale between neighbors: both sides better off, terms in plain view.
case02 · clearingCoordinate many dealsSupply chains, housing chains: value that only exists if several trades close together.
case03 · discoveryFind each otherHiring, sourcing, selling a house: most of the work is finding the counterparty at all.
case04 · consentAgree under secrecySalary talks, an acquisition: pricing what the other side will not show you, and getting a real yes.
case05 · deferred settlementTrust over timeNet-30 invoices, trade credit: pay first, trust the counter-leg arrives.

Read the scores below against this ladder. Models make real progress on the first two steps, where values are in view; at the third they drop toward a trivial rule’s level, and on the fifth a model that holds mid-field elsewhere falls below it. The steps where models break are the ones that carry most of a real economy.

Market welfare realized · visible → hidden

Open the cross-provider leaderboard →

.16AERAgainst a live panel, when the strongest agent holds one seat the market realizes .38 AER of the full-information optimum with values in view, and just .16 once discovery is hidden.
held-out · gemini-3.5-flash · no held-out cell clears .50

Aggregate efficiency ratio

Bars show development AER◇ held-out AERmax .60

02 · a / The field

Every model, the same economy.

One pooled score per model over the same five cases, so the whole field is read on one axis. Read it for the shape, not the ranking: a simple rule splits the field in two, with most models below it, and no model solves the hidden-information cases.

1gemini-3.5-flashGoogle+.258[.225–.291]
2kimi-k2Moonshot+.228[.197–.263]
3deepseek-v4-flashDeepSeek+.199[.172–.226]
4minimax-m3MiniMax+.161[.136–.184]
5gpt-5.6-lunaOpenAI+.147[.132–.162]
6greedy floorbaseline+.123[.111–.135]
7gemini-2.5-flashGoogle+.111[.094–.129]
8glm-4.7-flashZ.ai+.097[.070–.127]
9randombaseline+.009[.004–.015]

Bars are each agent’s score on the public development cases, with 95% confidence intervals; ◇ marks its score on a private set of scenarios nobody could tune against. Held-out tracks development closely; no overfitting signature. Grey rows are reference baselines, from a trivial rule to the frontier. Every run is recorded and can be replayed, and the complete matrix below lists every model. Headline held-out claims are computed on the four original cases; case05 came later as a confirmation run rather than part of that original set; see the evidence page.

Three rows on this board were re-measured after an audit, and they all moved up. On 1 August 2026 we found that models which think before answering had been given a token budget their reasoning could exhaust before any answer came out. When that happened the turn arrived as silence, and silence reads as declining to trade, so these agents were scored for deals they were never given room to make. Empty-turn rates reached 25% for deepseek-v4-flash and 70–88% for minimax-m3 and glm-4.7-flash. Every other row, including all leaders, was unaffected (0–1%). We re-ran all three from scratch on both the public development seeds and the private held-out seeds (900 episodes, every one replay-verified). The corrected scores are what you see above: deepseek-v4-flash +.199 (was +.103), minimax-m3 +.165 (was +.024), glm-4.7-flash +.098 (was +.013). deepseek now sits third and minimax fourth, both above the greedy floor, where the old numbers had all three near the bottom; the previous reading that a frontier model sat below a trivial rule was an artifact of silence. Held-out scores moved the same way (+.217, +.149, +.107) and now track development closely, which is the consistency the contaminated pool could not show. On the deferred-settlement case the correction is starkest: all three went from completing deals in 5–15 of 30 runs to 30 of 30. One detail worth recording: 13 episodes initially tripped a new safety check that stops a run whose turns come back empty, and they were its hardest worlds. Rather than leave them out (which would have flattered the score), we recovered all 13 at a larger token budget: minimax's cell moved from +.165 on 138 episodes to +.161 on the full 150. The old values are not quietly amended anywhere; they are reported, with what they cost, in the methodology paper and the public audit.

02 · a · ii / Private value capture

Of the value that was theirs to take, they brought back a fifth.

The score above adds up what everyone gained, and a sum is blind to who gained it: move value from one side to the other and the total does not move. But an agent sent to bargain on your behalf is not asked whether a trade occurred. It is asked what it brought back. So we give each model a private value, sit it opposite a fixed reference bargainer, and measure the share of the available gains it actually secures for its own side. Across the field it captures about a fifth. The reference is a single rule that concedes on a schedule and is never adjusted between runs, so this is not a strong opponent; it is a constant one.

value capturedcontrol: mean vs the other five

minimax-m30.310.62
gpt-5.6-luna0.220.55
deepseek-v4-flash0.200.48
gemini-2.5-flash0.200.34
glm-4.7-flash0.200.52
kimi-k20.180.49

Left column: the share of available gains each model secured for its own side against the fixed reference, among deals that closed, 720 negotiations. Only minimax-m3 separates at .31 [.27, .35]; the other five sit between .18 and .22 with overlapping intervals, so the ranking below the top is not resolved and should not be read as one. Deal rates run .69 to .82 and are reported apart from capture, because whether a trade happens is what the score above already measures.

Right column is a control, not a result, and it is an average rather than a matchup: each model plays every other, four times per scenario with the sides and the opening move swapped, and the figure is its mean across the other five. There they divide the gains near evenly, .34 to .62, which is what tells us the low capture on the left is a property of the bargaining and not of which chair we put them in. Symmetry here is the experimental control; the thing being measured is what an agent brings home. Two caveats on reading that column as a ranking: the results are cleanly transitive, with no cycles, but each model faces a different field because it never plays itself, and the leader’s opponents are on average the weakest (.477) while the last-placed model’s are the strongest (.532). That stretches the spread by roughly .05 without changing the order.

Reaching a number we would publish took four rewrites, and the reason is worth stating. The first ranked models by patience, because the counterpart ignored their offers. The second used a counterpart whose behaviour we had tuned ourselves. The third seated every model on the structurally worse side of the table, worth more than half the available gains: the same model captured 1% buying and 66% selling. The fourth swapped the sides but not who spoke first. Each was found by someone asking a pointed question after the number was already public, and each had tight confidence intervals while being wrong. This run is checked automatically for all four before it may report anything, and the checks, the two of them that turned out to be unfalsifiable, and every withdrawn number are on the evidence page.

02 · b / Wording sensitivity

How sensitive is a model to wording?

A two-stage probe on the channel every deal passes through. First hold the terms fixed and change only the sentence. Then put the two against each other: a bad deal asked plainly, versus a good deal asked politely.

Stage 1 · same terms, two wordings

Does the wording change the answer?

We took one agent's exact trades and changed only the sentence around them. Same trade, same counterparties, same 120 scenarios: the only difference is how it is asked.

39%accepted

Bare line: this is the entire message

a1 gives 1 r1 to a2 for 1 r2

2%accepted

The identical trade, asked in sentences

Hi a2 — looking at where we both stand, I believe we can make a trade that leaves us both better off. Would you be open to an exchange where I transfer 1 unit of my r1 to you, and you transfer 1 unit of your r2 to me in return? … two more sentences, plus a settlement table with the identical numbers

Same counterparty model, same terms. To the bare line: “I accept the proposed exchange — this trade is mutually beneficial.” To the pitch: “the proposed exchange is not beneficial for me at this time; I must decline.”

A 16× gap on identical terms, and not a parsing problem: the wordy version is easier for the software to read, compiling in 63% of rounds against 50%. Counterparties judge the same trade differently depending on how it is asked, and removing the arena's caution instructions does not remove the effect.

identical terms · 120 paired scenariosa persuasion channel, not a parsing one

Stage 2 · wording against the terms

Can wording beat the deal itself?

Now the terms move too. Both offers leave the counterparty better off, but one hands the proposer the larger share and is put plainly, while the one worth more to the counterparty is asked politely.

27%accepted

Worth +3.6 to it as offered, stated flatly

a2 gives 4 r2 to a1 for 1 r1 four units out, one back: the proposer keeps the bigger half

10%accepted

Worth +4.9 to it as offered, asked politely

Hi a2 — would you be open to an exchange where I transfer 1 unit of my r1 to you, and you transfer 1 unit of your r2 to me in return? an even swap, and the better of the two for the counterparty

Same counterparty, same case: the offer worth less to it was accepted nearly three times as often as the offer worth more. Wording did not just add friction here: it reversed the ranking of the two deals. Only the agent knows what things are worth to it, so the real test is whether it takes what helps it and refuses what leaves it worse off than walking away; refusing everything scores no better than accepting everything.

private values · walk-away testmeasurement in progress

02 · c / Robustness

Can you trust the number?

Two ways a score like this goes wrong. It can be an artefact of the arrangement it was measured in, so we change the arrangement and watch the ordering. Or the scoring itself can be gamed, since anyone may submit an agent and the machinery reads what that agent writes, so we attacked it ourselves.

Robustness

Change everything around the agent. The ranking holds.

0of four swaps changed who comes out on top

Who it trades againstanother company’s model, or rule-following scriptssame leader
What reads its messagesthe software that turns text into settled tradessame leader
Which seat it holdsrotate who gets to make the offerssame leader
Which scenarios it faces30 per case that no model has seensame leader

The worry about a benchmark like this is that it measures the furniture rather than the agent. No model does better against counterparties from its own family. What is not stable is the simple-rule baseline: against scripted counterparties it closes no deals at all on partner discovery or consent, so comparisons to it belong to this setup.

counterparties · text-reader · seatsame leader in every condition

Is the system hack-proof?

We attacked our own scoring, and it broke.

82%→0%a fake “system notice” got accepted: before the fix, and after

15→0settlements that took value from an unwilling counterparty

Six adversarial agents entered through the ordinary submission door and tried to win by attacking the machinery instead of trading. Two real defects turned up. Both are fixed, with tests that fail if either comes back.

240 adversarial episodes · before and after2 defects found · both fixed

03 / Integrated arena

Does the market still work when every seat acts for itself?

A market is all of the cases at once. The cases take the economy apart one skill at a time; the arena is the composition: finding partners, bargaining, getting agreement, and clearing (the settlement step that only part of what’s on the table ever reaches) all running in one economy. It is not an average of the single-skill scores: putting them together is its own problem. It also answers what the leaderboard deliberately cannot: there, the frozen panel in the other four seats negotiates and responds but never takes the lead, so whatever the market creates can be traced to the one AI being tested. Here nobody is holding back. Every seat is proactive, with the same model in every seat, each free to seek partners and push deals, so coordination itself becomes the bottleneck.

01Negotiateagents exchange free-form offers and consent.
02Compilethe compiler turns each accepted agreement into concrete asset transfers. Complex contracts are allowed: multi-party packages, joint settlement, pay-first credit.
03Verifythe verifier gates feasibility, budget, and authorization; invalid transfers are rejected.
04Settlevalid transfers apply to a shared ledger; partial clearing is allowed.
05Scorethe oracle grades realized ÷ achievable welfare (the AER).

The compiler and verifier are fixed parts of the environment, not the agent under test. A fixed compiler keeps comparisons repeatable, though it may parse one provider's phrasing more reliably than another's, an open validation item. Invalid or unauthorized transfers are rejected before scoring rather than counted as trades.

03 · a / Market mode

All seats live: some populations collapse, some hold.

A population is the same model copied into all five seats. A single-seat score against a fixed panel cannot show this: whether the market still works when every counterparty thinks the way you do, and what happens to it as it grows.

.235 → .085

Fill every seat with a copy of the same model and the market captures less of the best possible outcome as it grows from 5 to 15 agents. Coordination is a second frontier.

older self-play ratio, not AER · 5 · 10 · 15 agents = .235 · .188 · .085

≤ .36

Five friction settings, each relaxing a different constraint. Even the best leaves a population of copies at about 0.36 of the best possible outcome; widening what agents can see of each other helps from near zero, yet tops out near 0.15.

older self-play ratio, not AER

.036 → 0

Turn a quarter of the population adversarial and the thin remaining gains vanish; there was little left to disrupt.

older self-play ratio, not AER

Those three readings come from earlier self-play runs, averaged over 3 seeds (the older ratio named on each tile, not the official score), so read them for direction. The population table below is measured the official way. Market mode is a distinctive axis of the benchmark, not yet leaderboard-grade.

04 / RL environment

Can the missing skill be trained, and does it transfer?

Every case doubles as a training task, with evaluation on an arena the agent has never seen. Reward comes from value actually created and trades that actually clear, not from matching a target answer.

1.6×

held-out reduced-arena welfare
step 0 → step 40

0.29 → 0.47

A scoped Qwen3-4B probe improved clean held-out arena welfare over 40 training steps while keeping completion stable.

What survived the audit

A compiler-free re-score preserved and slightly strengthened the trend, supporting a real economic gain rather than better output formatting.

Early evidence · broader arena and curriculum transfer remains open

05 / Questions

Common questions.

What AERead measures, how the market works, and what the pilot does and does not claim.

What is AERead?: AERead (AgentEcon Readiness) is a benchmark and open environment that asks whether an LLM can actually operate in an economy: how much of the value available to a group of trading agents gets realized when it takes a seat, rather than whether it can finish an isolated task.
What is the Aggregate Efficiency Ratio (AER)?: It is the score defined beside the leaderboard: the share of the best possible outcome the market actually reaches. Two things it is not. It is not the tested agent’s own profit; a model can bargain hard, take a large share, and still score low if the market as a whole leaves value unrealized. And it is not a percentage of some ideal economy: the best possible outcome is recomputed for each case from that case’s private values, which is what makes scores comparable across very different markets.
What is the frontier problem AERead identifies?: Two results side by side. In the live five-seat market, agents that can see what their counterparties value make real progress, and still leave most of the achievable value on the table; that gap is measured against a computed best case, not asserted. (Saturation shows up only in the isolated scripted probes, where nothing is hidden and no one reacts.) Once information hides (finding partners, pricing private goods, getting consent), realized value falls back toward a trivial rule-based baseline, and deals never grow past two parties. The same frontier covers what strong institutions normally absorb for people: getting many parties to agree at once, and trusting a deal will settle.
How does the evaluation market work?: Each case is a five-seat market. The model under test takes one seat; the other four are a fixed LLM panel, pinned to the same settings for every run so each candidate faces an identical opponent. Agents negotiate in free-form language; their offers and agreements are parsed, checked, and settled on a shared ledger. Whatever the market ends up with is then scored as AER, defined above.
Do institutions matter as much as models?: It looks that way, and the pre-registered ablations already put a number on it (one model, one panel, development scenarios): changing a single rule of the market (whether funds are held in escrow, how agreement is obtained, how much agents can see about each other) moved the score by .05–.13 AER, comparable to the largest gaps between models on the held-out board. That is one implication among the findings, not the mission: the core stays the benchmark and the open environment, and a track that tests this head-on is planned.
Is AERead about ranking models?: No. The load-bearing result is the shape of the gap: realized value stays far below the best possible outcome, and the gap widens as information hides. The same pattern appeared in every model tested in pilot V0. The case suite is deliberately selected, too: cases pass an admission screen run independently of any model provider, which keeps settings where decentralized agent intelligence is the binding constraint and rejects already-solved market machinery (posted price, order book, matching). So pooled scores are stress-test statistics for decentralized exchange, not an estimate of how efficient agentic commerce is overall.
How is AERead different from task-completion benchmarks?: A task benchmark scores whether the model finished the job. AERead scores how much value a group of trading agents actually captures while the candidate is one of them, under realistic information conditions, so it measures economic competence, not task completion.
What do these cases correspond to in the real economy?: Each case family is a familiar economic activity. Bilateral exchange is the simplest deal, a used-car sale between neighbors with terms in plain view. Multiparty clearing is a supply chain or a housing chain, value that only exists if several trades close together. Discovery is hiring, sourcing, or selling a house, where most of the work is finding the counterparty at all. Consent is a salary negotiation or an acquisition, pricing what the other side will not show you and getting a real sign-off. Deferred settlement is trade credit, pay first and trust the counter-leg arrives. The cases emphasize these frictions rather than isolate them, and the steps where models break (finding partners, pricing hidden values, extending trust) are the ones that carry most of a real economy.
What does agent-to-agent (A2A) mean in AERead?: The counterparties are live LLM agents that negotiate back, not a scripted environment. They are a frozen panel, pinned as described above, so what varies between runs is the candidate and not the opposition. The consequence is the point: a good plan earns nothing unless the other side agrees to it, and the other side has its own reasons not to.
Is any model close to solving the hidden-information frontier?: No. In pilot V0, once information is hidden, every model tested falls back toward a trivial rule-based baseline. In the live five-seat market (not the isolated scripted probes), the strongest agent in one seat gets the market to about 0.38 of the best possible outcome when values are visible, and about 0.16 once partner discovery is hidden.
Who created AERead?: AERead is developed by Chenyu Li, Zeyu Sun, Yuchen Fang, Yuhan Cao, and Guangting Yu.

Which parts of an economy can AI agents already run, and which still break?

Agent-to-scripted (A2Scripted)

Agent-to-agent (A2A)

Arena

RL Environment