Tested: Replacing Prompts with 'Genomes' for LLM Agents
MonokoEloba · reddit · 2026-08-26
The author tested GenOS, an environment driving LLM agents via a versioned YAML 'genome' instead of massive prompts. By mutating traits (e.g., risktolerance) and 'breeding' specialized agents, the author achieved emergent TDD, bypassed RAG context limits, and avoided multi-agent 'ping-pong' loops.
Test Case: Fixing a PaymentProcessor.ts with 38 lint errors and a silent security hole (currency conversion bug).
Results:
- Prompt Baseline (Failed): Simple agent missed the bug; Expert agent (600 tokens rules) took 3 iterations and high token cost to fix.
- Genome Mutation (Success): With verificationthreshold = 0.80, the agent refused to touch production code directly, writing tests first (emergent TDD) to catch the bug. Adjusting syntaxstrictness fixed lint errors with zero extra tokens.
- Breeding vs. Swarms: Breeding 'SecurityAuditor' and 'PerfEngineer' genomes created a child agent that wrote an AES-256-GCM engine passing all checks at 21 ops/ms in a single pass, avoiding swarm debates.
More from coding & agent
- Salesforce launches 'Agent Script Decoded' video series on building agents — msrivastav13 · 2026-08-26
- LikeC4: Generate Live Software Architecture Diagrams Directly from Code — tom_doerr · 2026-08-26
- Study finds Agent harness impacts benchmark scores more than the model itself — rohanpaul_ai · 2026-08-26
- Why giants like Ramp and Spotify build their own AI infrastructure — bendee983 · 2026-08-26
- AI skills and agent magic enable scalable knowledge management — dSebastien · 2026-08-26
- Phonebook: An MCP-First Tool for Mobile UI Galleries — orelzion · 2026-08-26