Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSet
FactSet treats skills as versioned product features and the harness as their runtime. Routing descriptions, model-specific evals, access controls, and governance matter as libraries grow.
A minimal skill-capable harness needs **a registry, a system prompt, and file reading**; executable skills also need a sandbox. The registry exposes names, descriptions, and paths, while the agent progressively loads only the instructions needed for a task.
Write descriptions as distinct routing signals tied to user intent, and organize skills around jobs rather than internal data models. Rerun evals after model changes because skills are **contracts versioned to a model**, not passive documentation.
A minimal skill-capable harness needs **a registry, a system prompt, and file reading**; executable skills also need a sandbox. The registry exposes names, descriptions, and paths, while the agent progressively loads only the instructions needed for a task. Write descriptions as distinct routing signals tied to user intent, and organize skills around jobs rather than internal data models. Rerun evals after model changes because skills are **contracts versioned to a model**, not passive documentation. Putting every skill in the prompt stops scaling. The talk suggests shortlisting beyond roughly **10 skills** and adding hierarchy, metadata, ownership, lifecycle rules, audits, and tool allowlists when a library reaches hundreds.
This turns skills from Markdown conveniences into model-bound product contracts with routing, evaluation, ownership, security, and lifecycle requirements. It confirms progressive loading as the scaling mechanism, while narrowing simple skill-pack approaches: once libraries grow, descriptions must support reliable selection and executable skills require sandboxing, allowlists, audits, and reevaluation after model changes.