WTF is Agent Instruction Bloat?
Agent instruction files that grow long, stale and contradictory until they make the agent worse.
Also called: instruction bloat, agents.md bloat
TL;DR
Every time the agent slips up, someone adds a rule to AGENTS.md. A year later it's 900 lines, half outdated, some contradicting each other, and the model skims the middle anyway. More instructions stop meaning better behaviour.
How it works
Thoughtworks (Caution, Radar Vol. 35) recommends: turn instructions into deterministic checks where possible (linters, tests, hooks); re-evaluate instructions against each new model and delete ones no longer needed rather than only adding; and use progressive disclosure so only task-relevant instructions load. They also note hand-written instruction files outperform AI-generated ones. It is the cautionary flip side of harness engineering.
Origin
- Coined / introduced by
- Thoughtworks Technology Radar
- When
- Nov 2026
- Story
- Flagged as Caution in Radar Vol. 35.