Long-running agents fail by re-litigating old decisions, not by crashing
ClickOk5811 · reddit · 2026-11-22
A Reddit post breaks down a silent failure mode of multi-session agents: after context compaction, agents confidently propose approaches that were tried and rejected, because summaries keep state and facts but lose the reasoning behind rejected options. The author proposes a State/Facts/Decisions framework and lists open questions: where decisions should live, how to retire stale ones, how to keep logs from becoming noise, and how to verify agents actually honor them.
More from coding & agent
- Harrison Chase: buying Copilot won't transform teams without a knowledge factory — hwchase17 · 2026-11-22
- OpenAI's $200 Codex Edition ModRetro handheld turns prompts into playable cartridge games — shaunralston · 2026-11-22
- Aiden Bai doubles down: "** agents!!!" — aidenybai · 2026-11-22
- Cognition CEO Aiden Bai says he never reads the code, just watches agent test videos — aidenybai · 2026-11-22
- NBER study of 500k+ GitHub devs: coding agents boosted commits ~240% but releases only ~30% — Antique_Atmosphere97 · 2026-11-22
- Dev complains reviewing AI-written code means juggling Zed, GitHub, Cursor and Devin — pqoqubbw · 2026-11-22