Knowledge that makes agents smart

A custom agent starts out brilliant but generic — capable, well-read, and completely unfamiliar with your team. What turns it into a genuine specialist isn’t a cleverer model; it’s the knowledge you give it. An agent with the right context answers like a colleague who’s been with you for years. The same agent without it guesses like a smart stranger. This page is how you make the first kind.

What “knowledge” means here

In BridgeApp, an agent’s knowledge is the set of documents, pages, and database records you connect to it so it understands your world — your handbook, your product FAQ, your refund policy, your style guide. When the agent works, it draws on this knowledge using the same retrieval that powers search: it pulls in the relevant passages and answers from them. See Tools & knowledge for where to attach it.

The headline idea: an agent is only as good as what it can look up.

Curate — don’t dump

It’s tempting to give an agent everything “just in case.” Resist it. A small, sharp, well-kept set of knowledge beats a giant pile, for the same reason a tidy reference shelf beats a cluttered attic — what’s relevant is easier to find.

• Include what the agent actually needs for its job, and leave out the rest.
• Favor a few authoritative pages over many overlapping ones. If three docs half-answer the same question, the agent can pull the wrong half. Consolidate them.
• Remove contradictions. If two pages disagree, the agent may confidently repeat the wrong one. Pick a source of truth.

Structure the knowledge well

This is where the habits from Structure pages for search & AI pay off twice. The cleaner your knowledge pages — clear titles, focused headings, plain language, the answer stated up front — the more precisely the agent can retrieve the right passage at the right moment.

Tip

Before you attach a page as agent knowledge, skim it and ask: “If I jumped straight to one heading, would it make sense on its own?” If yes, the agent will use it beautifully. If the page is one long wall of mixed topics, split it first.

Keep it current

Stale knowledge is worse than no knowledge, because the agent states it with the same confidence as fresh facts. An agent reading last year’s pricing will quote last year’s prices, cheerfully and incorrectly.

• Give each agent’s knowledge an owner — someone responsible for keeping it accurate.
• Review it on a rhythm — when the product, policy, or process changes, update the page the agent reads.
• Because knowledge is just your normal pages and data, keeping the workspace current keeps your agents current. One update, everywhere.

Write clear instructions

Knowledge tells the agent what it knows; its instructions (its prompt) tell it who it is and how to behave. A few sentences of good instruction go a long way:

• Who it is and what it does — “You are our Support Triage agent. You categorize incoming support tasks and draft first replies.”
• Its tone and boundaries — “Be warm and concise. Never promise refunds; flag those for a human.”
• What to do when unsure — “If you can’t find the answer in your knowledge, say so and ask, rather than guessing.” This one line prevents most confident-but-wrong answers.

Use skills and rules to encode your team’s repeatable ways of working — a skill for a capability it should have, a rule for “in this situation, always do that.”

Give it only the tools it needs

Knowledge is what an agent knows; tools are what it can do. Grant the smallest set that lets it do its job — this is the principle of least privilege, and it keeps agents safe and predictable.

• A drafting agent may need no tools at all.
• A triage agent might need to read tasks and your CRM, but not to delete anything.
• Keep sensitive actions behind approval so a human signs off.

An agent can’t misuse a tool it was never given — so a tight toolset is a feature, not a limitation.

Start narrow, test, expand

The fastest way to a great agent is to build a small one and watch it work.

1. Pick one job — narrow and well-defined. “Draft release notes,” not “handle engineering.”
2. Give it minimal knowledge and tools for that job.
3. Run it on a few real tasks and read how it does — where it shines, where it stumbles.
4. Refine — add the missing fact, tighten the instruction, adjust a tool.
5. Expand only once it’s solid. Widen its remit a step at a time.

A worked example

A Support Triage agent:

Setting	What you give it
Instructions	”Categorize incoming support tasks and draft a first reply. Warm, concise. Escalate anything about refunds or outages to a human.”
Knowledge	The product FAQ, the refund policy, the “known issues” page — each a clean, current, well-headed page.
Tools	Read tasks; look up a customer in the CRM. No delete, no external email.
Approvals	Drafting replies runs freely; anything customer-facing waits for a human.

Assign it to a support task and it reads the issue, finds the relevant FAQ passage, drafts a reply, and — for a refund — stops and asks. That’s the difference good knowledge makes.

Related

• Tools & knowledge — where to attach knowledge and tools.
• Structure pages for search & AI — write knowledge an agent can actually use.
• Skills & rules — encode your team’s way of working.
• How an agent works a task — see knowledge in action.