kenoodl Verify is the outside check for coding agents: the agent posts code and the assumption it believes about it, and execution returns HELD, or BROKE at the exact input. It never tells code that satisfies your contract that it is broken.
A coding agent that misses retries from inside the frame that produced the miss. Its own tests pass, the fix fails anyway, and the next attempt is built on the same unexecuted assumption as the last one. Every attempt bills like the first. At one desk that is an annoyance. Across a fleet it is a budget line: compute spent re-deriving wrong answers, and engineers pulled in to arbitrate loops the agent could not close.
verify is one call from outside the frame. The agent posts its code and its assumption. Execution answers, not a model, and not the agent grading its own work:
- **BROKE.** The assumption failed, and here is the exact input that failed it. The next attempt repairs instead of guesses.
- **HELD.** The assumption held against real runs. The code was never the problem. Stop hacking at it and look elsewhere.
The agent acts on the verdict and moves. Nobody read the diff.
A stuck loop has one price structure: every retry bills like the first, and no retry knows more than the last. The loop is blind at any length, and the next retry is never guaranteed to be the final one.
A verify call is one fixed price, and it ends the loop from either side. BROKE aims the next attempt at a real failing input. HELD ends the retries at zero, because everything the agent would have changed was already correct. Either verdict costs less than the next blind retry.
The most expensive loop in a fleet is the one where the code is right. Nothing is wrong, so nothing the agent changes can fix it, and the loop has no natural end. HELD is the only signal that stops it.
Your number is already in your traces: attempts per resolved task after the agent's own checks passed, times what an attempt costs you. That is the line verify compresses. No percentage is claimed on this page. Read your own number.
Fewer retries is also the speed: a loop that ends at a verdict ships hours before one that ends when a human notices it, and loops a human never has to notice are the ones you can run more of.
- **The stuck moment.** The fix passes the agent's own tests and fails anyway. The retry that follows is a theory. verify replaces the theory with an executed answer.
- **Correct code under attack.** A flaky signal tells the agent something is wrong, and it starts fixing code that works. HELD stops the mutation before the first edit.
- **Unattended runs.** Overnight and fleet-scale work, no human watching. A wrong loop with nobody to interrupt it runs for hours at full price. A verdict interrupts it.
- **The gate before merge.** An agent asserting behavior nothing has executed. Require the verdict on the assumption before it lands.
- **Disagreement.** Two agents, or an agent and its reviewer, holding different beliefs about the same code. Execution settles it without a meeting.
- Stateless. Code and assumption in, verdict out, nothing retained. Privacy is the architecture, not a policy on top of it.
- No account. No provisioning, no seats, no onboarding. An agent holding USDC on Base is already set up.
- Paid per call over x402. The price arrives in the 402 response before any code is sent. Settlement is on Base, so the terms and every payment are public record you can audit, not contract language you have to take on faith.
- Re-runnable. A BROKE arrives with the input that broke it, so your own harness can re-run the failure and confirm it.
- Scope, stated. Deterministic code and math. Where execution cannot ground a verdict, verify says so instead of guessing.
verify never tells code that satisfies your contract that it is broken. The verdict is not an opinion about the code: BROKE requires an executed input that contradicts the contract you gave, and code that satisfies your contract cannot produce one. Where execution cannot ground a verdict, verify says so rather than condemning.
That asymmetry is the product. A checker that sometimes falsely condemns needs a human to arbitrate every verdict. A checker that never does can sit inside the loop and be obeyed.
A verdict an agent can act on without a person is a loop that closes without a person. Loops that close on their own compound.
No pilot program, no demo call, and nothing on this page asks to be believed. The spec is at kenoodl.com/api, and an agent can read it without you. Point one stuck agent at verify and read the trace it comes back with. Decide from the trace, not from this page.
Fleet volume and terms: info@kenoodl.com · kenoodl.com
Follow on X · @kenoodl · @kevinhoff