Should you post-train a model for your agent harness? Dev seeks real cost data
Small_Luck8177 · reddit · 2026-09-06
A Reddit developer runs an agent that uses the same tools on the same tasks with frontier models, and is considering post-training an internal model to be better and cheaper in their harness.
They're soliciting feedback from two groups:
- Those who tried it: engineer-weeks and GPU spend required, and how quickly harness changes forced a retrain?
- Those who considered it and passed: was it opportunity cost, or low expected value?
The thread is a useful case study in whether fixed-tool, fixed-task agents justify private fine-tuned models.
More from coding & agent
- Spec-to-APK entirely on an Android phone via Termux CLI and AI — Digital_Otorongo · 2026-09-06
- Astra review: executes hour-long agentic tasks while iterating on the plan without losing the thread — ivan_bezdomny · 2026-09-06
- Stop picking coding models on vibes: run your own A/B/C refactor tests — robleclerc · 2026-09-06
- Orca, the 62k-star ADE for running parallel coding agents, recommended for Mac Mini dev setups — kieranklaassen · 2026-09-06
- Dev Abandons LLM Agent Harness After 883 Commits and 8 Months, Open-Sources the Wreckage — dancingwithlies · 2026-09-06
- Dev one-shots a UEVR fork with GPT 6 Astra for true stereoscopic VR windows — pvncher · 2026-09-06