Agent Economic Readiness

Benchmarking agents’ readiness for value delivery rather than just task completion.

Benchmarks like Terminal-Bench and GDPval evaluate whether AI can complete a technical task or produce professional-quality work. AERead evaluates whether AI can turn information, resources, and delegated authority into beneficial economic outcomes through interaction with other actors. We read AI adoption differently: not as the automation of existing job functions, but as a way to raise the value delivered in critical economic activities.

01

Why this?

Take an M&A deal. Investment banking is among the higher-value occupations GDPval covers, yet the repetitive junior work that fills those hours, the models, decks and data rooms, carries only a small share of what the deal is worth.

When competent work products are not enough, what prevents AI agents from realizing available economic value?

02

What we do

An open-source environment for running economic benchmarks, with selected cases in deal structuring and contract negotiation, for agent builders and research teams. Agents get a mandate rather than an assignment, and are scored on the outcome their decisions produce.

03

How we do it

Run agents in simulated economic tasks. Settle only what is feasible and authorized. Check outcomes with objective verifiers, and keep every run replayable from its records. Rubrics reviewed by human experts, for assessing how an agent got there, are in development.

A passed test or a graded deliverable is not a settled outcome.

We measure realized surplus, principal outcomes and constraint adherence, not merely the quality of a recommendation or the completion of a transaction. The target column is where the benchmark is headed.

Code and documents still matter here. The difference is whether the artifact is where evaluation ends, or an instrument that moves an outcome.
Dimension Task completione.g. Terminal-Bench Professional outpute.g. GDPval Economic outcomeAERead’s target
Starting pointA task instruction in a sandboxed terminalAn assignment to produce specified workAn economic situation plus a mandate
The agent mustCarry the task through in that environmentProduce a useful deliverableDecide on a beneficial course of action and carry it out
Other partiesNone; the environment only runs commandsMostly described in the supplied contextActors who can withhold information, refuse, counteroffer, or fail to perform
What changesFiles and system state in the sandboxThe work productAllocations, contracts, payments, obligations, exposure to risk
SuccessThe task’s tests passOutput quality, as experts judge itEconomic outcomes, within authority and other constraints
Typical failureTask unresolved, tests failingIncorrect or unusable workMissed surplus, harmful commitments, failed coordination, losses to the principal

Case study, October 2026

Agents that do not stop to check

A buyer can pay a familiar supplier’s premium or check cheaper listings first. Every decision is priced against the exact best policy given what the buyer knows. On 100 held-out worlds, Claude Opus 5.5 acting directly committed at once in 64% of episodes. Asking it for a valued plan first recovered about 80% of the loss.

Average loss per episode against the best available policy, by condition
  1. Directbought at once in 64% of episodes$84.72
  2. Plan firstbought at once in 28%$18.34
  3. Plan suppliedbought at once in 34%$17.16
  4. Plan and values suppliedbought at once in 2%$3.93

Claude Opus 5.5 at low effort, about 99 episodes per condition on 100 held-out worlds screened so that checking pays. One model, and a loss in simulation, not yet a measured business outcome.

We are building a community of industry reviewers, engineers and researchers to uncover overlooked scenarios and review tests.

Join as a reviewer or contributor