FL-TR-2026-01 · preliminary
Software Factory Technical Report
Software development with an unattended agent loop.
We ran 14 coding agents for roughly 4.5 hours on a private production codebase. Each worked in a separate lane. An automated gate checked commits; model review, a blind spot-check, and operator sign-off preceded integration and deployment.
- Factory commits
- 132 written / 110 initially integrated
- Token accounting
- 848M, including cached reads / ≈$493 at recorded list prices
- Estimated value
- 77 non-minor bug-fix commits / ≈98 engineer-hours, AI-estimated
- Regression comparison
- 181 baseline Python failures / 174 candidate failures / zero new failures
One run, one codebase, no control group. Value estimates are model judgments, not measured savings. Costs are a frozen snapshot. Subsequent fixes and verification are recorded separately in the paper.
Files
The download contains the factory harness, prompts, synthetic tests, report builders, figures, accounting, and sanitized review records. Application source and private Git history are excluded.
Code: MIT. Paper and data: CC BY 4.0. The arithmetic and factory tests can be reproduced; the private application's evaluation cannot.
Citation
Infected Research / FrackLabs. 2026. Software Factory Technical Report: Software development with an unattended agent loop. FL-TR-2026-01.