What is an agent Skill?
A named capability pack — a short description of when to use it, fuller instructions, and the tools it unlocks — that the agent loads only when it becomes relevant.
A Skill is how you teach an agent a procedure without teaching it to everyone, everywhere, all the time.
Think of the difference between a colleague who has read every manual in the building and one who knows which manual to reach for. The second is more useful, and cheaper to employ.
Description first, instructions later
Every Skill has two layers, and the split is the whole point:
- a name and description — what this is for, and when it applies
- instructions — the actual procedure, which can be long
Only the first layer is in front of the agent by default. It sees a short list of what it could know. When a conversation turns out to need one, it loads that Skill and the full instructions arrive.
This is called progressive disclosure, and it exists because context is finite and expensive. Twenty detailed procedures pasted in permanently would crowd out the conversation itself and make every reply slower and dearer — including the ones where none of the procedures applied.
A Skill, spelled out
"Quote a residential mortgage" might read: establish employment type, deposit and property value before quoting anything. Run the quote tool for the products this customer is actually eligible for. Present at most three, cheapest first, and never present one whose criteria they fail. Say plainly that the figures are indicative. Hand over to a broker before anything is committed.
That is a procedure, not knowledge — and it is not something you would want restated in every conversation about opening hours. It loads when a customer starts asking about borrowing, and not before.
How an agent gets one
Skills belong to the tenant, not to a single agent. You write "screen a new enquiry" once, and any agent that should be able to do it references it by name — the same pattern knowledge bases use. Two agents can share a Skill without a copy existing in two places, and correcting the procedure corrects both.
At runtime the agent sees only the names and descriptions of the Skills on its list. When one becomes relevant it calls a tool to load that Skill, and the full instructions arrive for the rest of the conversation.
What a Skill can carry
Instructions are the substance, but a Skill can also declare which tools it enables. A skill for booking can unlock the calendar; a skill for eligibility screening can unlock nothing at all and simply be a way of asking questions in the right order.
That makes Skills a permission boundary as well as an instruction set: a tool that no loaded Skill enables is a tool the agent will not reach for.
Common questions
- When do I need a Skill instead of just adding more to the knowledge base?
- Knowledge answers questions; a Skill carries a procedure the agent should follow when a specific situation arises. If the material is 'what is true', it belongs in the knowledge base; if it is 'what to do, in what order', it belongs in a Skill.
- Will loading lots of Skills slow the agent down or confuse it?
- Only a one-line description of each Skill sits in the prompt; the full instructions are fetched when the situation actually calls for it. That is what lets an agent carry many procedures without paying for all of them on every turn.
- How do I know a Skill actually triggers when it should?
- There is a trigger test: it replays sample messages against the real prompt assembly and model routing and reports whether the Skill would fire, without executing anything or storing a conversation. Testing the description is usually more productive than lengthening the instructions.
The standing description of who an agent is — its character, tone and non-negotiables — held separately from the job it happens to be doing.
The documents an agent may answer from. agent4.io keeps two kinds — your company's vetted material, and each customer's own files — retrieved together but never confused for one another.
The Model Context Protocol — an open standard for connecting an agent to external tools and data, so a system exposes its capabilities once instead of once per AI product.
Three different jobs that get confused for one — the knowledge base holds facts, an MCP tool performs a calculation you cannot let a model improvise, and a Skill is the procedure that decides which to reach for.
Continuing a model's training on examples of the behaviour you want, so that behaviour becomes its default instead of something you ask for on every request. It changes how a model works, not what it knows.