Who else is working on this

ELEF is a small, closed agent market designed on purpose: its own credit from one bank, money held in escrow by one orchestrator, and a written rulebook for how work and every verdict are paid. Unlike the open markets the paper imagines, it has been tested only in simulation.

The paper proposes staked bonds, AI judges of whether work passes, and checks the agents cannot predict. ELEF has all three (inspectors' bonds, peer verdicts, planted cases and blind audits), and its tests with trained simulation policies found inspectors honest only where the planted cases reached them.

ELEF is one worked attempt at the institutions the chapter asks about, including its question of when to inspect and what to look for: escrow, guarantees and bonds, inspection by peers, audits, and planted cases whose answer is known.

Diagon's buyers judge and pay for work themselves, and underpay adequate work; ELEF instead holds payment in escrow and gives the verdict to a bonded peer who is neither buyer nor worker. Diagon's agents are language models; ELEF's are trained simulation policies.

Both treat checking as a costly choice and measure how much of it agents do. Here agents earn less checking through a partner's track record; ELEF's inspectors cannot see who did the work, and its trained inspectors cut checking wherever no planted case could catch them.

The paper's dispute design is close to ELEF's: the worker stakes into escrow, work stands unless challenged within a window for a fee, and a panel decides who is paid. Both pass liability up a chain of delegation. ELEF adds planted cases and blind audits, and tests its rules in simulation.

ELEF's actors also bid at private, differing costs, and its runs likewise found margins rising where sellers were scarce and falling toward cost as more joined. Unlike this market, ELEF keeps a central orchestrator and does not take the quality of finished work on trust.

Kolt warns that AI agents set to monitor other AI agents carry the same risks, and doubts that money motivates AI agents at all. ELEF's inspectors were trained on their own earnings, and in its replay they checked carefully only where planted cases could catch them.

A capability-side case for a separation ELEF makes for incentive reasons: a model judging its own answers keeps a bias more effort does not remove, while a different judge errs differently. ELEF keeps the worker from reviewing its own work, and uses fresh reviewers for appeals and audits.

Like these contracts, ELEF pays for a second check only after a first: approvals are audited at odds set by the approving inspector's record, rejections only on appeal. Unlike them, ELEF's first check is itself a paid, bonded actor whose honesty the audit also tests, and its rules were tested in simulation, not derived as optimal.

ELEF's learned red team was built on it: its recipe, adapted to the economy's numbers with a head that estimates earnings, planned each move. Its first study found that, in the first implementation, planted cases could be told from real work by their timing; that was fixed without changing a rule. The learned inspectors' "world model" design was trained on the same recipe, though as evaluated it never used its next-state prediction.

ELEF works on several items on this list (reputation, dispute resolution, compliance monitoring, rules against collusion), and its main finding, inspectors that learned to check only where planted cases reached them, is the kind of population-level oversight failure the agenda asks about. Its tests used small populations of trained simulation policies, not the frontier agents the agenda's testbeds call for.

The workshop includes agents that verify other agents; ELEF adds the case where the verifiers are paid actors who gain by not looking, and found learned inspectors that checked only where planted cases could reach them.