Decagon shares 19+ ablations on using GEPA for test-driven prompt optimization in production
kastnerkyle · x · 2026-09-12
Key takeaways
Decagon researcher Roy Wang details how the team used GEPA (Reflective Prompt Evolution) to optimize prompts for a production classification task.
- GEPA background: introduced at ICLR 2026 as a gradient-free optimizer built on DSPy, using natural-language reflection instead of policy gradients; reportedly beats RL methods like GRPO by up to 20% with 35× fewer model rollouts.
- Use case: a supervisor model that analyzes conversations and produces structured judgments with reasoning traces, serving as a last line of defense against hallucinated outputs.
- Method: a test-driven approach to prompt engineering, backed by 19+ ablation experiments.
- The post covers three findings that challenged conventional wisdom about prompt optimization, arguing configuration details matter for production reliability, not just benchmark scores.
More from coding & agent
- OpenAI launches Agents API in public beta: Codex harness as a cloud agent service — minchoi · 2026-09-12
- Hermes Agent open-source project hits 3,000 contributors — Teknium · 2026-09-12
- Pro tip: append .md to any Apple HIG link to get an AI-friendly Markdown file — rudrank · 2026-09-12
- Data science career path shifts from ML training to building LLM apps and agents — mdancho84 · 2026-09-12
- The AI-native SDLC: agents across planning, coding, testing and deployment — Pavan_Belagatti · 2026-09-12
- Litho (deepwiki-rs): generate C4 architecture docs from code, ADRs and SQL schemas — techNmak · 2026-09-12