Knowledge base, tools or Skill — which does what
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.
Most of the trouble people have configuring an agent comes from putting something in the wrong one of these three. So here is the dividing line, with a mortgage broker as the example.
The three questions, and where each is answered
"What are the terms on your fixed-rate product, and would I qualify?"
This is a fact. It is written down somewhere in your material, it is the same answer for everyone who asks, and the only correct source is your own documents. → Knowledge base.
Retrieval finds the passage about that product and the agent answers from it, citing what it used. If your documents do not cover it, the correct outcome is "I don't have that", not a plausible paraphrase.
"What would my repayments be on 380,000 over 25 years, given my income and this deposit?"
This is a calculation. It has one right answer, the rules behind it are long and fiddly, and in a regulated business a wrong number is not an embarrassment but a liability. → MCP tool.
You do not want a language model doing arithmetic on affordability rules by improvisation, however capable it is. You want it calling the same function your own systems call, getting a number back, and reporting that number. What is auditable afterwards is a tool call with inputs and outputs, not a paragraph of reasoning that happened to arrive somewhere.
"Given everything you've told me, which of these products should I actually be looking at?"
This is a procedure. It needs several things to happen in order, with judgement about what to ask next. → Skill.
The Skill holds the instructions: establish employment type and deposit before quoting anything, run the quote tool for the products the customer is eligible for, present at most three, never present a product whose criteria they fail, and hand over to a human before anything is committed. It calls the MCP tool as part of doing its job, and it can lean on the knowledge base for the terms it explains along the way.
The fourth line people miss
"Facts" is really two things here, and getting them mixed up is the most expensive mistake on this page. agent4.io keeps a tenant knowledge base — your rules, prices and regulations, vetted and authoritative — separate from each customer's own knowledge base, holding the documents they uploaded.
The mortgage question needs both: your lending criteria come from yours, their payslips come from theirs, and the answer is the comparison. They are retrieved together and injected as separate blocks, one framed as policy to follow and one as the customer's own material for reference — so the agent never quotes a customer's own document back at them as if it were your policy. See knowledge base.
The rule of thumb
| If it is… | It belongs in… |
|---|---|
| Written down and the same for everyone | The knowledge base |
| Computed, and must be exactly right | An MCP tool |
| A sequence of steps with judgement in between | A Skill |
Or, more bluntly: facts are retrieved, calculations are called, procedures are followed.
The same three in another business
An immigration practice draws the lines in exactly the same places:
- Knowledge base — the documents required for each route, current processing times, your fee schedule
- MCP tool — the points calculation for a skilled-worker application, run against the official criteria rather than remembered
- Skill — "screen a new enquiry": ask status, timeline, employment and prior refusals in the right order, run the points tool, and produce a structured profile for the caseworker with the routes worth discussing
Why an agent carries several Skills
One agent can hold many — screening, quoting, chasing documents, booking — and it loads only the one the conversation turns out to need. A quoting Skill can unlock the quote tool; a screening Skill can unlock nothing at all and simply be a way of asking questions in the right order.
That is also a permission boundary: a tool that no loaded Skill enables is a tool the agent will not reach for. Which is the real reason to encapsulate the sensitive calculation as a tool in the first place — not merely so it is correct, but so that reaching for it is a decision you configured rather than one the model made.
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.
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.
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.
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.