Cisco открыла FAPO: автооптимизация LLM-пайплайнов с диагностикой на каждом шаге

Cisco Foundation AI has released FAPO to the public — a system for automatic prompt optimization that operates at the level of an entire pipeline, not individual requests. Claude Code handles the orchestration, transforming the tuning process into a fully autonomous cycle.

Why

One Prompt Is Not Enough Most modern AI products are not a single request to a language model, but a chain of interconnected steps: one extracts data, a second classifies, a third reasons, a fourth generates the final answer. When overall quality drops, manually finding which exact step loses accuracy is a lengthy and unclear task. Existing prompt optimization systems, including the popular GEPA, treat each prompt in isolation: they don't account for how changing one step affects the inputs and quality of subsequent ones. FAPO is built on fundamentally different principles — it sees the pipeline as a whole and optimizes it as an interconnected system.

How the

Optimization Cycle Works FAPO operates in four stages that repeat iteratively: Diagnosis: the system runs the pipeline on a test set and measures accuracy at each step. This identifies the specific step where quality is lost — the authors call this step-level failure attribution. Variant Generation: for the problematic step, FAPO proposes fixes at three levels — prompt (new instruction wording), parameters (temperature, top-p), pipeline structure (add or remove a step).

Independent Validation: each variant is evaluated by a separate LLM-reviewer agent that makes decisions based solely on metrics — without bias from the change author. Iteration: the cycle repeats until target accuracy is achieved or the attempt budget is exhausted. Claude Code serves as the orchestrator: it sequentially runs diagnostics, invokes variant generators, passes data to the reviewer, and applies approved changes.

All of this happens without human involvement.

Benchmarks: FAPO vs GEPA

Cisco compared FAPO to GEPA — one of the leading methods for automatic prompt optimization. The result was convincing: FAPO won in 15 out of 18 comparisons across model + task combinations.

"The approach with error attribution at the step level is particularly

effective where the problem is not localized in a single prompt, but is spread across multiple transitions in the chain," the authors note in the technical documentation. FAPO's key advantage is precisely in the multi-step context: GEPA and similar systems look at each prompt separately. FAPO understands that the output of one step is the input to the next, and optimizes the chain accounting for these dependencies. This is critical for agentic systems, where changing an early step cascades through all subsequent ones.

What

This Means Cisco has published FAPO under an open license — the system can be deployed on your pipelines right now. For teams building multi-step AI agents, it's a way to automate what previously took weeks of manual trial-and-error. The choice of Claude Code as the orchestration engine is a telling signal: corporate AI tooling is increasingly being built on the Claude ecosystem.