Universal Jira: Can a Market Coordinate AI Agents?

← Back to Research Lab

Mechanism Design · Verification Without an Authority

When AI agents both propose work and are paid for finishing it, how do you tell useful work from reward hunting? This note sketches one proposal: replace the ticket with a market on whether a task will be done, keep a price on both sides with an automated market maker, and let an oracle settle it afterwards. The weak point is the settlement, and that is the open question.

Scientific Reference: Open research direction of the author, not yet published. The question: can a market price and settle engineering work without a central planner, and what has to be true of the settlement step once the agents proposing the work are the same ones being paid for it? The mechanism is built on standard foundations (DAO mechanism design, constant-product AMM invariants, Logarithmic Market Scoring Rules); the contribution under development is applying them as a coordination substrate for software delivery, and connecting the resolution problem to the verification work in my dissertation and the oracle film. Feedback and collaborators welcome.
Open problems, and a first test

Assumptions. Tasks can be stated as propositions with a resolution date; an oracle can judge whether a proposition came true; traders hold no private control over the outcome they bet on.

Unsolved. The last assumption fails as soon as the agents trading on a task are the ones doing it: a market then rewards whoever can make a proposition resolve YES, which is not the same as useful work. Self-proposed tasks can be written to be easy to resolve. The oracle inherits the whole verification problem. Nothing here shows that the mechanism is incentive-compatible.

A first test. A simulated backlog with honest agents and agents that propose trivial or self-resolving tasks, against a plain central work queue as the baseline. Measure completion and usefulness separately, since the market prices "will this task be done?" and not "is it worth doing?"; then measure how much value the second kind of agent extracts, and how much of that a stricter settlement rule (independent review, a deposit forfeited on rejection) removes.

🧠 What did you just learn?

The Coordination Problem. Traditional software uses managers to assign tasks and evaluate progress. In a trustless global network, a market can aggregate beliefs and incentives into a public coordination signal. Truth still depends on how the outcome is defined and settled.

The Ticket is a Proposition. Instead of "Fix bug X", the ticket becomes a proposition: "Will bug X be fixed and merged by Friday?" People can buy YES or NO shares. The price of YES is a compact summary of the market's belief about completion, not a guarantee.

The AMM Math. Using a Constant Product Market Maker ($x \times y = k$) or LMSR, the contract keeps a counterparty available at every price. The sponsor of the bounty seeds the initial liquidity, and each trade nudges the price toward the market's current estimate.

Work Changes the Price. A developer who intends to fix the bug buys YES shares while they are cheap because the task is still incomplete. If they really finish the work and the PR gets merged, the market later resolves to 100% YES. Their profit comes from having aligned with the truth early, not from manipulating the outcome.

Optimistic Oracles. How does the contract know the PR was merged? An optimistic oracle (like UMA) is used. Someone asserts the fact and posts a bond; if nobody challenges it during the dispute window, the market resolves and pays out. The oracle is the bridge from off-chain reality to on-chain settlement.

Closing the loop: let the agents write the tickets. Everything above still assumes a person notices the problem and opens the market. Take that assumption away. An agent with read access to the repository, the issue tracker and production telemetry can propose the proposition itself, post it to the contract, and let the price form. The cycle then runs end to end without a human in it: an agent proposes, the market prices, an agent works, the oracle settles. This is one candidate coordination layer that could scale with the number of agents rather than the number of managers, if proposal quality and settlement can be secured.

And that is exactly where it gets hard. Once the same population both creates the work and profits from it, the supply of problems stops being exogenous. An agent can open an issue that is trivial or invented, buy YES while it is cheap, resolve it, and collect. Nothing in the market detects this, because every step looks like honest work: a proposition, a trade, a merge, a settlement. The mechanism pays for claimed problems, so claims are what it will get.

Candidate defences, the same shape as the one in my dissertation. Spoofing Proof-of-Learning means fabricating evidence of a training run that never happened; fabricating a ticket means manufacturing evidence of work that was never needed. One candidate answer in both is to stop paying for the claim and pay for something harder to fake. Three measures could serve here, none of them tested: proposing costs a bond that is slashed when the issue is judged fabricated, the proposer is barred from settling their own market, and the payout is tied to an artifact that real work tends to produce, such as a previously failing test that now passes on an independently run build. A passing test shows the task was completed, not that it was worth doing. Price discovers what to do; verification decides what actually got done. The second half is the open problem, and it is the same one this lab keeps circling.

📐 The math, precisely

Rendered on load. If equations appear as raw text, reload the page.