loading…
Package reusable domain capability and load it only when the task needs it.
A skill packages the instructions, knowledge, tools, examples, scripts, and validation needed for a recurring capability. It gives an agent a stable way to perform domain work without loading every domain into every request.
| Skill element | Purpose |
|---|---|
| Metadata | Name, description, triggers, owner, version, permissions |
| Instructions | Procedure, boundaries, and escalation rules |
| Knowledge | Domain references loaded when needed |
| Tools / scripts | Deterministic operations and external actions |
| Examples and evals | Expected behavior and regression coverage |
At startup, the agent sees a lightweight skill catalog rather than every detail. It selects a relevant skill, loads its instructions and resources, executes it, and releases unnecessary context. This reduces context rot and keeps ownership modular.
A customer asks: "I want to tile my shower. What do I need, and can it arrive Tuesday?" One monolithic assistant would need deep expertise in construction, merchandising, and fulfillment. Skills distribute ownership:
Project guidance skill (trades team) → Materials-list skill (merchandising) → Delivery-window skill (fulfillment)
Each team maintains the knowledge and deterministic logic it already owns. The agent composes the skills for the customer's task without the central AI team becoming expert in every department.