FAPO: Fully Autonomous Prompt Optimization of Multi-Step LLM Pipelines

FAPO (Fully Autonomous Prompt Optimization) is a new framework that automatically optimizes multi-step LLM pipelines by iteratively refining prompts and, when necessary, restructuring the pipeline architecture itself. The system demonstrates significant performance improvements across multiple benchmarks, achieving up to 33.8 percentage point gains over existing optimization methods.

FAPO addresses a critical limitation in current LLM optimization approaches: the assumption that prompt engineering alone can resolve performance bottlenecks in complex, multi-step pipelines. Traditional prompt optimization tools typically ignore how failures propagate through chains of retrieval, reasoning, and formatting operations, missing systemic architectural issues. This framework represents a meaningful advancement in automated pipeline engineering by combining diagnostic capabilities with constrained structural modifications.

The research builds on growing recognition that LLM systems often fail not because individual steps underperform, but because their interactions create compounding inefficiencies. FAPO's approach of exhaustively evaluating intermediate outputs before proposing changes provides transparency absent in black-box optimization methods. The use of Claude Code as the optimization agent itself demonstrates the recursive potential of AI-assisted engineering—using language models to improve language model systems.

For enterprises deploying complex LLM applications, this work has immediate practical implications. Organizations currently relying on manual prompt tuning or limited optimization tools could potentially unlock substantial performance gains, particularly on security-critical tasks where the framework shows robust improvements across multiple model sizes. The 14.1 percentage point average gain over existing baselines suggests real competitive advantage for early adopters.

The framework's effectiveness on security benchmarks (CVE-to-CWE classification) indicates broader applicability beyond traditional NLP tasks. Future developments likely involve expanding the scope of structural modifications permitted, reducing the human annotation required for evaluation, and adapting FAPO for real-time optimization in production environments.

• →FAPO beats existing optimization baseline GEPA in 83% of tested configurations with mean gains of 14.1 percentage points.
• →The framework intelligently escalates from prompt-only optimization to structural pipeline changes when bottlenecks are identified in the chain logic.
• →Security task performance improved significantly, with accuracy gains of 2.0-7.1 percentage points on CVE classification depending on model size.
• →FAPO diagnoses failure modes by inspecting intermediate pipeline steps rather than treating the entire system as a black box.
• →Automated LLM pipeline optimization represents an emerging category of AI tooling that could reshape how enterprises develop and deploy language model applications.