If submitting is cheap, what should a bounty actually pay for?

Bounties used to self-filter. Submitting one was expensive: a person had to read the issue, learn the codebase, and write the patch. That cost did the triage for you.

Coding agents remove that filter. Submitting is now close to free, which means a bounty with a fuzzy acceptance criterion quietly turns into a machine for generating plausible-looking submissions that someone still has to read. You have not bought work at that point, you have bought review load, and you are paying for it with the scarcest thing you have.

Four things I think have to change, and I would like to hear where people disagree:

Acceptance has to be mechanical. If a human must read the diff to decide whether the bounty is met, it is priced wrong. A command that exits zero, a benchmark that moves, a failing test that now passes.

Pay for the verified outcome, never the submission. Obvious, still the most common mistake. Any scheme that rewards attempts rewards volume.

Cap concurrent claims. Without a claim mechanism you get twenty parallel attempts at one task and nineteen closures nobody budgeted for.

Price the reviewer's time into the task. If it takes forty minutes to review, that is part of what it costs, and it should push you toward work with cheap acceptance checks.

The honest consequence is that the set of tasks this fits is narrower than people want. It works for flaky test fixes, dependency bumps with a green suite, coverage on a named module, mechanical migrations. It does not work where correct is a judgement call, where a security boundary is involved, or where the hard part is deciding what to build. Dressing those up as bounties just relocates the cost onto the reviewer.

I wrote the longer version up here: https://sonaarajyan03.github.io/agent-handoff-notes/bounties.html and there is a more careful treatment of the verification and reward side at https://wagglet.com/blog/gamification-ai-work-bounties-rewards (disclosure: I work on Wagglet).

Has anyone actually run a bounty program since agents got good? I am curious whether the submission flood is real in practice or whether I am solving a problem that has not arrived yet.