SoduWork: open-source agent workbench adds AI review + human approval to verify agent output
soloboy4567 · reddit · 2026-11-20
The team open-sourced SoduWork, a self-hostable, MIT-licensed agent workbench (public beta) aimed at making task verification less dependent on an agent simply reporting that it finished.
The workflow has three layers:
- Predefined acceptance criteria as hard constraints;
- A separate AI review step where another model checks output against the criteria;
- Final human approval as the backstop.
The team is upfront that they haven't measured a correctness improvement: the second model may repeat the same mistakes or confidently approve wrong results. They're asking two open questions — when both agents may share similar failure modes, how useful is independent model-based review, and how should such a setup be evaluated?
It's an engineering exploration, not a new ML algorithm or benchmark, and they want feedback from people working on agent evaluation and automated verification.
More from coding & agent
- A 2026 company is just a folder: every department as .md agent files — mdancho84 · 2026-11-20
- Google AI Studio can now build Android apps from a prompt and install them privately, no Play Store listing needed — TawohAwa · 2026-11-20
- No task tracker fits agentic dev: 10-100 concurrent tasks need new patterns — bigblueboo · 2026-11-20
- SelfGraph: open-source versioned knowledge graphs for AI agents over any graph DB — JeremyCMorgan · 2026-11-20
- Why one researcher's immunology model was the code he'd only ever write by hand — iskander · 2026-11-20
- IBM engineer Tejas Kumar launches 3-week live course on AI agent evals and harnesses — TejasKumar_ · 2026-11-20