optimize-agentic-workload

Reduce latency and cost in agentic workloads by routing policy models through a verifiers harness to compare quality, latency, and cost against fixed tool sets. This structured evaluation enables model A/B routing and staged GEPA interventions to find cheaper, faster solutions.

Run a model A/B evaluation by establishing a baseline harness with fixed tools, snapshotting tool outputs, and comparing different policy models on quality, latency, and cost metrics. This ensures fair comparisons across multi-turn, tool-calling loops.

You can optimize both read-only searches and state-mutating API workflows by applying cost- and latency-reducing strategies across these distinct paths. The skill enforces deterministic harnessing and safety gates before optimizing state-mutating operations.

GEPA intervention is a staged optimization path for agent policy models that follows baseline capture and model A/B routing. It provides a structured approach to reducing workload costs while maintaining quality through gated handoffs and deterministic harnessing.

Snapshotting tool outputs is required to establish deterministic harnessing and safety gates before optimization. This baseline capture ensures fair quality, latency, and cost comparisons when routing different policy models against the fixed tool set.

Use a fixed tool set when the policy model slots vary in multi-turn, tool-calling loops. Keeping tools fixed enables fair comparisons of quality, latency, and cost across different models during A/B evaluation and staged interventions.