Cisco’s new FAPO framework uses Claude agents to autonomously optimize complex multi-step LLM pipelines.
On June 19, 2026, Cisco AI introduced Fully Autonomous Prompt Optimization (FAPO), an optimization system driven by Claude Code agents to transition LLM pipelines from baseline prompts to target accuracies. Built as a multi-tenant evaluation framework, FAPO utilizes isolated directories to run unrelated tasks side-by-side without interference. Its domain-agnostic core engine, hephaestus, handles evaluation, scoring, and LangGraph-defined state graphs across providers like OpenAI, Baseten, and SageMaker.
Users supply an initial prompt and a dataset. FAPO splits the data into a validation set and a test set. Claude Code then orchestrates a closed six-stage loop: Evaluate, Attribute, Propose, Review, Compare, and Iterate.
FAPO tackles errors across three escalating levels—prompt edits, parameter changes, and structural changes (like altering chain topology)—exhausting one level before advancing. Step attribution classifies failures into four categories: prompt-addressable (format and reasoning) or structural-addressable (retrieval and cascading). To prevent overfitting, guardrails isolate aggregate data, keeping variants immutable under an independent reviewer.
In performance benchmarks, FAPO outperformed GEPA in 15 of 18 model comparisons, yielding a +14.1pp mean gain. On complex pipelines like HoVer and IFBench, where it escalated to structural modifications, it won all six pairs with a +33.8pp mean gain. GEPA led only on AIME by a narrow 3.1pp. FAPO excels at multi-step use cases, such as multi-hop QA, classification, and optimizing ReAct agents via tool-calling trajectories.
