@gsemetfr@zachlloydtweets Makes sense, the input really varies. We keep a story as the anchor because each change gets checked against it at the end. Does anything check the result against the folder?
@zachlloydtweets Agreed on measurement. We record review rounds and cost for every merged story, so a change to the setup has something to be compared against.
People keep asking what a software factory actually is. For us it's the same written-down steps for every story. A person reviews the story and the approach before any code, and every change gets checked against its spec after.
@clay@sophiebits We moved that conversation earlier. A person goes over the story and the technical approach with the agent before any code gets written, so the big PR shows up with the design already agreed.
@arncof@zachlloydtweets That works as long as you're the one watching every run. It breaks when you're busy or someone else joins, because the checks live in your head. A factory writes them down, so every change gets the same spec check whether anyone's watching or not.
@gsemetfr@zachlloydtweets Agreed, the front end is where most setups are thin. Ours starts with a story review and a technical review, where a person goes over the story and the approach with the agent before any code gets written.
@jacewbuilds@zachlloydtweets We give the repair loop a budget. When it runs out, the run stops and a person gets asked. We'd rather pull someone in early than let the agent keep guessing.
@zachlloydtweets Agree on workflows in code. We'd add a defined way out of the loop. Each change gets checked against its spec and repaired up to a budget, and then a person gets asked. Otherwise it loops forever or ships whatever it has.
@thsottiaux The cloud computer lack real utility given it's small disk size (30GB). Strongly recommend an option to expand it - even if it cost a higher % of usage.
Moving up a layer changes what you judge: not the agent’s first draft, but the loop around it.
A factory reviews every story against its spec and repairs it until it passes, or a person is asked. Without that exit, it’s a retry loop with a bill.
We record rounds and cost.
Agree. “Factory��� is not a perfect term but the move up in abstraction is real. It’s a shift towards engineering agentic systems away from directly prompting agents.
@dextroxin@agentroar@robocurve Yes, I agree. If you have the ability to interfere with the video feed going to a local model then the battle is already lost.
@stevenwhatevr Genuinely curious about the 50-100 a day. Is that branches ready for review, or merged PRs? From the README, numbrone itself opens no PR and runs no tests, so what sits between “worker says done” and merge: CI, a human read, both? And what’s the revert rate?
@surlymochan Independent is key. Green CI only proves the code agrees with its own tests. Passing means it meets the spec and the planner’s contract: scope, interfaces, acceptance criteria. The reviewer holds the diff to that contract. Mutants check the tests; the contract checks the code.
Your agent's tests pass and the code is still wrong, because the same model wrote both. Plowline 0.3 mutates only the lines a change touched and fails the train if no test notices. One gate line: builtin:mutants. Free, local, five engines. https://t.co/3QP6DnPi1d
@hwchase17 Async is the right call. The next snag we hit with long parallel tasks: by the time one finishes, main has moved, so we retest its branch merged with everything that landed before calling it done.
@yeemio Good list. One I'd add next to the evangelist: a check every task has to pass (tests, a build, a screenshot), so trust comes from evidence rather than from whoever ran the demo.
@DrewWakefield5 Not just you. Two open issues on https://t.co/dQEkHv9G1R report the same fast drain this week (59398 and 59487). Adding your numbers there gives the team a case to reproduce.
@addyosmani One addition to checking the slices for overlap: two branches can touch different files and still break once merged, like one renaming a function the other calls. The check that catches it runs on the merged result before anything lands.
@solopribuilds Not dumb! Your computer may still be writing data even after a copy looks done. Ejecting flushes pending writes and closes access. Pulling it early can corrupt files or the drive’s file system. If it’s truly idle, it’s usually fine, but ejecting is safer.