Foreman open-sourced: a fast Jev model supervises coding agents, scoring completion and test sufficiency
udmrzn · x · 2026-09-18
Developer JoshARosen open-sourced Foreman, a "software factory foreman" built on TypeSafe AI's Jev model.
- Architecture: slow coding agents (e.g. Codex workers) do the actual engineering on the "factory floor," while Foreman sits above them with a fast decision model, continuously and independently assessing their work.
- Assessment signals: implementation completeness, requirements satisfaction, test sufficiency, whether the worker is stuck, and whether verification or human input is needed — output as confidence-scored signals (e.g. implementationcomplete .91, readytofinish .21).
- Decisions: based on these scores it chooses continue / stop / retry / verify / finish instead of blindly accepting agent output.
Positioned as an architectural experiment, the code is on GitHub (161 stars). A useful layered pattern for teams building multi-agent orchestration with automated acceptance of agent work.
More from coding & agent
- Paying a FasTrak toll invoice automatically with a Grok bot and Link — jeff_weinstein · 2026-09-18
- Honeycomb CTO on AI slop: half the company hates it, half hates the holdouts — mipsytipsy · 2026-09-18
- Credential-free MCP server template lets agents call tools without touching API keys — uzi24- · 2026-09-18
- Claude Handles 80% of My Sales Workflow — But It Can't Close the Loop — Prestigious_Rub5 · 2026-09-18
- Dev loses a day of benchmarks to Claude Opus 5, begs for Opus 4.5 back — julianharris · 2026-09-18
- Why only foundational model labs can self-improve with scaffolds, and why OpenAI/Anthropic likely already do — menhguin · 2026-09-18