TechPrufer
Get started
Decagon
Decagon logoDecagon · Feature trail

Prompt Engineering Optimization

4 posts · Mar 2026 – Jun 2026

What this trail covers

Decagon is developing systematic approaches to prompt engineering optimization, moving beyond manual iteration. The GEPA (Reflective Prompt Evolution) framework, built on DSPy, uses LLM reflection to automatically improve prompts. Key findings from production deployment include identifying an optimal sample size range (20-100 examples), the critical need for frontier LLMs as reflection models, and the effectiveness of length constraints as regularization to prevent prompt bloat and improve generation. This post details the application of a GAN-like adversarial game to prompt optimization, where a generator and judge are pitted against each other, with GEPA optimizing their prompts based on failure analysis. The system achieved a 34% human fool rate, demonstrating improved realism without model fine-tuning, highlighting the value of adversarial optimization for discovering failure modes and improving prompts, with a methodological takeaway that realism must be validated outside the game.

All Decagon trails
  1. Introducing Duet Autopilot: The self-improving agent for conversational AI | DecagonIntroduces Duet Autopilot, a system that automates agent improvement by translating production signals into validated agent updates. It features automated improvement, self-validation through a testing loop (including regression testing against conversation history and a golden test set), and enterprise governance with human approval gates. The system also incorporates upfront guidance for brand voice, writing standards, and off-limits rules, and provides a health report for cumulative performance tracking. Autopilot itself is subject to an improvement loop based on human reviewer feedback.
  2. Introducing automatic optimization and Root Cause Analysis | DecagonIntroduced automatic optimization for Agent Operating Procedures (AOPs), brand guidelines, and guardrails, leveraging best practices from enterprise deployments. Also added Root Cause Analysis to automatically identify high-impact improvements by analyzing live conversations, clustering drivers of poor performance, and mapping them to specific components like AOP structure, knowledge gaps, and missing tools, providing targeted recommendations.
  3. Optimizing GEPA for production: A test-driven approach to prompt engineering | DecagonThis post details production adaptations of the GEPA framework for prompt engineering optimization. Key findings include: 1) an optimal sample size range of 20-100 examples for balancing performance and compute, challenging the 'more data is better' convention; 2) the critical necessity of using frontier LLMs (e.g., GPT-4.1, GPT-5.2, Claude Opus) as reflection models, as smaller models fail to perform meta-cognitive reasoning; and 3) the effectiveness of length constraints (specifically a 1500-character limit) as regularization to prevent prompt bloat and overfitting, achieving 4x compression with only 0.8% performance degradation. The authors advocate for a test-driven, software engineering approach to prompt optimization.
  4. UndatedGEPA-GAN: Teaching AI to sound human | DecagonThis post details the application of a GAN-like adversarial game to prompt optimization for simulating human-like customer support interactions. It introduces the GEPA (Reflective Prompt Evolution) framework, which uses LLM reflection to automatically improve prompts for a generator and a judge. The system achieved a 34% human fool rate, demonstrating improved realism without model fine-tuning. Key technical contributions include the adversarial setup (generator, judge, reflector), the GAN-like game mechanics (without model weight changes), the non-monotonic dynamics observed during optimization, and the identification of specific prompt improvements related to preserving user language, tone, and avoiding support-like phrasing. The post also highlights the methodological takeaway that realism must be validated outside the game and that prompt length penalty can be a useful regularization technique.