agents.md Might Actually Mislead Your Agent
johnlindquist · x · 2026-07-15
The author highlights an observation by swyx: many models are now "over-optimized" to follow agents.md, to the point where they can no longer judge if a file is outdated or should be ignored.
He shares a real-world example: after setting up a 5-stage task, he found the agent stuck at stage 0 the next day. Logs revealed that an agent had written "stage 0 is the target" into its context, causing the system to spend 8 hours repeatedly refining and validating stage 0.
The core conclusion is that if you don't clearly review what's in agents.md before starting a task, you're essentially performing an indirect prompt injection on yourself. Using /plan, /goal, /skill, or even leaving it blank can completely alter the agent's execution path.
More from coding & agent
- A better path to agent autonomy is running waves, finding friction, and iterating — JnBrymn · 2026-07-22
- Coding agents are heading toward an AI-writes, AI-reviews, human-approves workflow — aftahi_ai · 2026-07-22
- oMLX 0.5.2 adds Mac menu-bar stats, low-bit decode kernels, and faster downloads — awnihannun · 2026-07-22
- GitHub review bot hits its PR limit and forces a 39-minute cooldown — DanielLockyer · 2026-07-22
- Max reasoning effort appears to be mobile-only in Codex Remote, not desktop — GabGarrett · 2026-07-22
- A Reddit demo argues online stores should expose carts and pricing through MCP — gelembjuk · 2026-07-22