All productsProduct studioShovon Saha
All products
Interactive prototype · Sample offers · No live checkout

These guides cover the standalone app. This demo does not create accounts, connect models or process purchases. Join the interest list to share your use case.

agentmesh
Try product demoHow it worksInterest listDocsPricing
Join
Documentation · browse topics
Building agents

Agents, knowledge and tools

An agent is three things: how it should behave, what it knows, and what it's allowed to do.

1 · How it should behave

The instructions box is just a note to your assistant, written in normal English. Say who it's helping, what to prioritise, and what never to do.

A good starting instruction
You help small studios buy camera gear.
Always check the catalog before searching the web.
Prefer sellers who ship within 5 days.
Never recommend anything over the buyer's stated budget.
Answer in three short bullets, then a one-line recommendation.

2 · What it knows

A knowledge base is a folder of your own text - price lists, policies, spec sheets, past answers. When a question overlaps with a document, the relevant part is handed to the assistant before it answers.

  1. 1Create a knowledge base and give it a clear name (e.g. “Return policy 2026”).
  2. 2Paste documents in. Plain text works best; one topic per document.
  3. 3Attach the knowledge base to your agent.
Rule of thumb
Ten short, focused documents beat one long one. The assistant picks documents by overlap with the question, so headings and specific words help.

3 · What it's allowed to do

Tools are the actions your agent can take. Platform tools are built in; API tools are yours.

  • product_search - look through the marketplace catalog.
  • web_search - read the public web.
  • research - gather several sources and summarise them.
  • list_orders - read the signed-in buyer's recent order history.
  • Write, payment, and messaging tools are withheld from ordinary model runs. Those actions use separate confirmed application flows.

A runtime API tool points at a public HTTPS endpoint you already run. Read-only GET tools can be used by an agent; private-network URLs and write methods are blocked from model execution. Use {{placeholders}} in the URL and the assistant fills them in.

Example API tool
name:        stock_check
description: Check live stock for a SKU in a warehouse.
method:      GET
url:         https://api.your-shop.com/stock?sku={{sku}}&wh={{warehouse}}
headers:     { "Authorization": "Bearer server-managed-token" }
You type
“Do we still have SKU-4410 in Rotterdam?”
What happens
The assistant calls stock_check with sku=SKU-4410 and warehouse=rotterdam, then answers with the number it got back.

Testing

Every run records the tools it called, the arguments, the results and the time each one took. If an answer is wrong, open the run and look at the step that produced the bad data - it's almost always a tool, not the model.