Dev builds adaptive harness that auto-scales LLM reasoning effort up and down
milpster · reddit · 2026-09-06
A Reddit developer proposes an alternative to manually fixing thinking-effort levels: modify your harness so the system adaptively raises or lowers reasoning effort between low and xhigh based on success and failure streaks.
The posted logs show it downshifts on "stable successful streaks" and escalates on "meaningful failures", combined with phase tracking (EXPLORE/DEBUG/IMPLEMENT). Subjectively it speeds up problem solving considerably, but no reliable quality evaluation yet — the author is considering a GPQA Diamond before/after test.
More from coding & agent
- Microsoft's FrontierTour: AI governance becomes everyone's problem as agents take action — donasarkar · 2026-10-23
- Grady Booch revives Weinberg's 1971 quip: software so buggy one woodpecker could destroy civilization — Grady_Booch · 2026-10-23
- One Agent Replaces Half an SDR Team: Full Lead-Research-to-Email Workflow — tibo_maker · 2026-10-23
- Why don't coding agent harnesses ship with their own sandbox? Only amp does — Aryvyo · 2026-10-23
- Dev builds a VST synth with Codex and MCP, lets the AI jam in his studio — Yamapama · 2026-10-23
- AgentsInTheCloud update: use your Claude subscription with built-in coding agents — lucasmeijer · 2026-10-23