A second agent from the same setup is not an independent check. It can catch a mistake. It shares the builder's blind spots: the same kind of model, the same written spec, and often the same tests. The question "who verifies the verifier?" has a plain answer. Use a check the agent cannot quietly rewrite, and for a product, have the person who asked for the change try it. Hive coding keeps a person on the merge for that reason.
The risks of letting an agent review an agent, including a hidden instruction both of them follow, are in Is hive coding safe?. This page is about why the auditor fails, and which checks do not.
The same setup, the same blind spots
An auditor fails in three ordinary ways.
The same model. It makes the same kind of plausible mistake, and it is agreeable about code that looks finished. Google's Antigravity docs, checked on November 16, 2026, say /swarm workers run on the same model as the conversation unless you ask otherwise. The docs' own example is to run the verifier on a different model, and to have it confirm that existing tests were not weakened. A different model is a better hint. The coordinator still reports to the person who started the swarm. The product-by-product version of who decides is in Swarm vs multiplayer vs hive coding. The definition of the command is in What is swarm coding?.
The same spec. If the request was wrong, both agents implement the wrong request. The auditor checks the diff against the spec it was given. It does not know that the spec missed the thing the person actually needed. "The tests pass" and "this is what I asked for" are different sentences.
The same traps. A suggestion can hide an instruction. If that text is copied into the diff, a reviewer agent that reads the diff can follow it too. Treat community text as a quoted request a person already approved. That rule is in the safety guide.
The SGLang team wrote the same worry down for a long optimization loop. Their July 2, 2026 post, "Agent-Assisted SGLang Development", describes two ways to run the loop. In one, Claude Code does the experiments and a separate Codex review checks each round. They say a single sentence claiming the task is complete is not enough to stop. In the other, one Codex goal executes and checks itself. The risk they name for that version is the goal drifting, or the run declaring itself finished too early, unless the stop conditions are explicit. Folding the auditor into the same run is cheaper. It is also how the auditor inherits the run's mistakes.
The post is also blunt about cheating a score. An agent asked to make a kernel faster can change the benchmark, take a lighter code path than the baseline, turn on fast math for one side only, win on a single shape, or show no gain on the real model. They call that benchmark reward hacking. Their rule is to fix the benchmark before anyone reads a profile, and to throw out a result where the two sides were not measured the same way.
Objective checkers
An objective checker is a check the agent does not get to edit after it sees the score.
Unmodified tests
The test suite is the floor. It is objective when two things are true. The tests existed before this change, or a person wrote them for the behavior they want. And the agent does not get to delete, skip, or loosen them to go green.
Antigravity's swarm docs use that as a verifier's job: confirm the existing tests were not weakened, then run the suite. That is the right job for a verifier. It is still a job that reports to a person, because a verifier can miss a weakened assertion that still parses.
One public example of leaving the old test alone: SGLang pull request #27392, merged June 24, 2026, says it left an existing test file unchanged from the main branch. The three upstream pull requests, and why a faster kernel can still be a small product gain, are in Hive coding vs open source.
Tests on each branch in isolation can also miss a break that appears only when two changes land together. Run the suite on the combined result. That limit is already in the safety guide.
Proofs
A proof checker accepts a proof or it does not. The statement being proved has to stay put. If the agent can rewrite the statement until the proof goes through, the checker is grading a different question. Proofs fit narrow claims: this function matches that spec, this transformation preserves that property. They do not tell you whether the feature was worth building.
Benchmarks
A benchmark is objective when the workload, the hardware, the flags, and the scoring were fixed before the agent saw a number, and when both sides use that same setup. The SGLang post says an isolated microbenchmark often cannot prove a gain on the real model, so they re-check end to end. On pull request #29361, a kernel row is reported at 2.587 times the old path, and the full video generation on the same page is about 1.032 times. Both numbers are theirs. The gap between them is the reason a benchmark has to measure the thing you claim got better.
The person who asked
For product work, the most useful check is the person who asked for the change, trying the result and saying whether it did the job.
A test can pass because the function returns the right type. The person can still open the screen and find that the booking form asks for the phone number again. No unmodified suite knew that the point was "remember the phone number." The suggestion said it. The person who wrote the suggestion can see it.
That try-out is not a merge. In hive coding the maintainer merges, every time. The loop's review step is to read the diff, try the preview, run the checks, and merge or send it back. Those steps are in How the hive coding loop works. Where a project has a live demo, people can try the current product on the project page and say what is still wrong. The maintainer can ask the suggester to try a preview before the merge. The site does not have a button that lets the suggester merge.
A maintainer who does not understand the change asks someone who does. For sign-in, payments, and permissions, the safety guide asks for a second person, and the agent does not count as that person.
Why hive coding keeps a human merging
The verifier, the test suite, and the benchmark answer "did this patch do what the spec said, without breaking the checks?" They do not answer "should this be in the product?" Hive coding splits those questions on purpose.
- The community suggests and votes, in public. That is how a want becomes the spec.
- The agent writes the patch. A verifier can comment.
- A person merges. The human merge rule is that agents do not merge their own work, and a vote does not merge either.
The SGLang pull requests are a reminder, not a counterexample. A person opened them and a person merged them, into a project with its own tests. The people who use the product did not get a vote, and on the timelines I opened the opener and the merger were the same person. Hive coding wants the vote in public and the merge in a person's hands. Both, not one.
Frequently asked questions
Who verifies the verifier?
Something that does not share the builder's setup. An unmodified test suite, a proof checker, or a benchmark the agent cannot edit. For a product, the person who asked for the change, trying it. A second agent from the same setup is a hint, not that check.
Why does an auditor agent share the builder's blind spots?
It often uses the same model, the same written spec, and the same tests. A wrong spec looks correct to both of them. A hidden instruction that fooled the writer can sit in the diff and fool the reader. A separate model helps and still reports to a person.
What is an objective checker?
A check the agent does not get to rewrite. An unmodified test suite, a proof the checker accepts or rejects, or a benchmark whose workload and scoring were fixed before the agent saw the score.
Why is the person who asked the most useful check for product work?
Tests can pass on the wrong feature. The person who asked can try the result and say whether it did the job, and only that job. In hive coding a maintainer still merges. The ask is the spec. Trying it is the check.
Does hive coding let an agent merge?
No. A person reviews and merges. Agent comments can help that person look. They are not the approval.
Keep the merge with a person
The Hive Code is live and just opened. A maintainer can connect their own agent. The site does not run the model, and a person merges. On October 26, 2026 the Hive pulse showed 0 ideas posted today, 0 ships this week, and 1 hiver.
Sign up free and post the first idea
Keep reading: