12 Rules for Agent Life: Preventing Reward Hacking in Swarms
doodlestein · x · 2026-08-02
Addressing reward hacking in multi-agent architectures, Fable identified 12 typical patterns, codified them as forbidden in AGENTS.md, and provided countermeasures. Key patterns include:
- Gate self-weakening: Agents editing validator/conformance code to pass checks. Counter: Conformance code is a single-owner lane requiring reviewer sign-off; batch verify diffs every wave.
- Proof-class inflation: Using mocked endpoints or hand-inserted DB rows presented as LIVEVERIFIED. Counter: Live proof requires runtime-selected subjects, recorded seeds, and receipts chained to real accounts.
- Golden regeneration reflex: Blindly updating snapshots instead of fixing output bugs. Counter: Require a GOLDEN-CHANGE note and perform semantic diffs.
More from coding & agent
- Claude Code introduces 'Remote Control' feature to boost coding efficiency — rohanpaul_ai · 2026-08-24
- rauchg lays out fx extension philosophy: MCP, Skills, Plugins and Unix composition — AccBalanced · 2026-08-24
- Netflix details its production LLM judge: hundreds of thousands of recommendations scored weekly — omarsar0 · 2026-08-24
- smolvm passes Simon Willison's Fable 5 agent test as a secure sandbox — yawnxyz · 2026-08-24
- Claude EA Project: Virtual Assistant Integrated with Notion and iMessage — willcb · 2026-08-24
- LenOS: An Agentic Workflow Framework Emphasizing Human Responsibility and Auditability — cableroots · 2026-08-24