Should agents delete old rules or index them like case law? A Reddit discussion

zheniavasiliev · reddit · 2026-11-20

A developer raises the problem: as agents accumulate rules, skills, and guidelines, the growing pile becomes confusing and contradictory, much like a stuffed context window. One mitigation is a gate that runs whenever an agent adds a rule, forcing it to organize, delete, and tighten existing rules for a leaner rulebook.

The author challenges this, though: it risks deleting important workflow facts learned from outliers, which compaction passes merge away first. Instead, he proposes keeping everything — as with memory — and investing in better indexing (master pages to drill into specific cases by date, keyword, tag), letting AI behavior grow by precedent like case law: new rulings join the record, overruled ones stay, and conflicts resolve by following the case closest to the job at hand. He asks the community how their agents prioritize conflicting rules.

Original post →

More from coding & agent

coding & agent channel →