Hey Product Hunt, Marvin from the RunAI Coder team.

We built this because we kept hitting the same wall with coding assistants: they
produce suggestions, and we produce everything after that. The unit of work we
actually wanted was a merged PR, so that is what the product delivers. You give
it a goal in plain English, it plans, fans the work out to parallel workers, runs
the tests, and files a delivery with the diff, the test run and the full trace.

The part we care about most is the boring part: "done" has to be earned. A goal is
marked complete only after a requirement-by-requirement audit against real
evidence, failing checks block completion, and approval gates stop it from
charging past a decision that is yours to make.

Receipts, since you should ask for them. One week of dogfooding, one developer,
one machine: 968 merged squash PRs, 98.8% first-pass (1,786 of 1,807 deliveries),
and 98 merged PRs on the best single day.

On cost, three stacked levers get a US$10 cash top-up to 8-9 deliveries a week,
and they are different kinds of claim. The provider's prompt-cache discount is
theirs, not ours; our only job there is not breaking our own prefixes. Our context
compression is ours and it is measured: 2.83x median, n = 7,222 production
requests. Batch and flex scheduling at half price is a price-list fact, not an
engineering achievement. The factor we put our name on is the 5.66x we actually
built, compression times saver, and billing is official API list price x1.15 with
the surcharge waived during early access, so budget for it coming back. Formulas
and sample windows: https://run.ceo/coder/perf?lang=en

What it is not. 21 of those 1,807 deliveries failed first pass, and we count them.
The compression number is a median: p10 is 2.26x, and our fleet-wide medians on
two channels are lower still at 2.45x and 1.89x, so plenty of workloads save much
less than the headline. Batch pricing is asynchronous, so interactive sessions do
not get the half price. The A/B that says compression costs no quality is 10 + 18
SWE-bench pairs and claims parity within noise, not superiority. And big refactors
still want a human reading the diff, which is why sign-off is a gate and not a
formality. We publish the engineering with the caveats attached:
https://github.com/RunAI-Coder/a...

Free starter pack is $25-30 of value, one-time, 14 days, no card.

One question back at you: what is the smallest piece of work you would hand to an
agent unsupervised? Those answers have moved our roadmap more than any feature
request.

Hey Product Hunt, Marvin from the RunAI Coder team. We built this because we kept hitting the same wall with coding assistants: they produce suggestions, and we produce everything after that. The unit of work we actually wanted was a merged PR, so that is what the product delivers. You give it a goal in plain English, it plans, fans the work out to parallel workers, runs the tests, and files a delivery with the diff, the test run and the full trace. The part we care about most is the boring part: "done" has to be earned. A goal is marked complete only after a requirement-by-requirement audit against real evidence, failing checks block completion, and approval gates stop it from charging past a decision that is yours to make. Receipts, since you should ask for them. One week of dogfooding, one developer, one machine: 968 merged squash PRs, 98.8% first-pass (1,786 of 1,807 deliveries), and 98 merged PRs on the best single day. On cost, three stacked levers get a US$10 cash top-up to 8-9 deliveries a week, and they are different kinds of claim. The provider's prompt-cache discount is theirs, not ours; our only job there is not breaking our own prefixes. Our context compression is ours and it is measured: 2.83x median, n = 7,222 production requests. Batch and flex scheduling at half price is a price-list fact, not an engineering achievement. The factor we put our name on is the 5.66x we actually built, compression times saver, and billing is official API list price x1.15 with the surcharge waived during early access, so budget for it coming back. Formulas and sample windows: https://run.ceo/coder/perf?lang=en What it is not. 21 of those 1,807 deliveries failed first pass, and we count them. The compression number is a median: p10 is 2.26x, and our fleet-wide medians on two channels are lower still at 2.45x and 1.89x, so plenty of workloads save much less than the headline. Batch pricing is asynchronous, so interactive sessions do not get the half price. The A/B that says compression costs no quality is 10 + 18 SWE-bench pairs and claims parity within noise, not superiority. And big refactors still want a human reading the diff, which is why sign-off is a gate and not a formality. We publish the engineering with the caveats attached: https://github.com/RunAI-Coder/a... Free starter pack is $25-30 of value, one-time, 14 days, no card. One question back at you: what is the smallest piece of work you would hand to an agent unsupervised? Those answers have moved our roadmap more than any feature request.

RunAI Coder

Autonomous coding agent: one sentence in, one merged PR out

Autonomous coding agent: one sentence in, one merged PR out