Machines / SF-01
MODEL SF-01
Software factory
Tickets in. Reviewed pull requests out.
A line of narrow agents that turns routine engineering tasks into tested, reviewable changes on an existing codebase — so engineers spend their time on the work that needs them.
In plain words: AI coding agents for your backlog.
INPUTA ticket from your project tool
OUTPUTA tested pull request, waiting for review
RUNS ONYour repo, your CI, your choice of model
OPERATORA person reviews, merges, handles escalations
Operating manual
Rev. 1 · SF-01§1
What it does
Reads a ticket, finds the code it touches, makes the change in an isolated branch, proves it with tests and a visual diff, and opens a pull request with everything a reviewer needs.
§2
Install when
- Small, clear tasks keep piling up
- The codebase is large or old, so context is the real cost
- There are tests you can lean on — or you'll add them
§3
Do not install when
- The work is architecture, not tasks
- Product decisions are still open
- Nothing can verify a change automatically
§4 · The line
Each station is a separate step with one job and its own check. A failure is caught where it happens, not at the end.
01TaskA ticket fires a webhook
02PlannerFinds files, writes a plan
03CoderEdits an isolated branch
04VerifierTests + visual diff
05Pull requestSummary, screenshots, task link
↺ Verifier fails → back to the Coder · limited retries → then a person, with the full trail
§5 · Safety features
Never pushes to mainEvery change goes through review.
Scoped permissionsAgents touch only the files and commands they're cleared for.
Retry limit, then a humanNo endless loops burning tokens.
Black-box recorderEvery run logged and replayable — failures become tests.
§6 · Gauges
What's watched on every install. Calibrated per team.Merge ratePRs merged without human edits
Cycle timeTicket to reviewable PR
Escalation rateHow often a person steps in, and why
Cost per ticketModel and compute spend per change
§7 · Parts bin
LLM — Claude / GPT / open models
Agent orchestrator
CI — GitHub Actions
Visual diff — browser runner
Project-tool webhook