'The End of Prompt Engineering': 300 Kimi K3 agents, one AGENTS.md, zero escapes in 41 days
JohnAlexander · x · 2026-09-13
A paper titled "The End of Prompt Engineering" argues that telling agents what to do is finished as a skill: prompts are wishes, but boundaries written in a file are law — versioned, owned, and read by every agent before every action.
The author validates the thesis with first-hand practice: one AGENTS.md file constraining 300 Kimi K3 agents with zero escapes over 41 days, inspired by a line in the K3 blog. Humans become boundary architects who define what never happens and audit the asks.
More from coding & agent
- px0: a read-only IDE for the agent era — <1ms boot, 16MB RAM, one Go binary — arpit_bhayani · 2026-09-13
- Dev ditched Claude for a browser-based Strawberry workflow under 35 seconds — damienghader · 2026-09-13
- An AI assistant inside a 20,000-note vault: systems first, AI second — dSebastien · 2026-09-13
- Building a meeting-notes-to-tasks Skill in Claude: hands-on walkthrough with skill-creator — eyishazyer · 2026-09-13
- AllSpark releases Iris-mini and Iris-pro, top open-weight search agents in their class — The Decoder · 2026-09-13
- stop-slop skill rewrites 37 passages and makes the draft longer — on purpose — bibryam · 2026-09-13