A custom agent is something you build to do a job. Where Kai answers questions from your content, an agent follows instructions you wrote and can act — look something up, change a record, write a report, ask a colleague.
Every agent is made of the same five parts. Understanding them makes the rest of this collection easier to follow — the other articles each go deeper on one piece at a time.
The text that tells the agent what it is and how to work. This is the part that most determines whether an agent is good, and the part people spend the least time on. An agent with vague instructions behaves vaguely no matter how capable the model is.
Each agent picks its own, independently of Kai and of other agents. A tagging agent that runs thousands of times a day and an agent that writes a weekly summary have very different needs, and nothing forces them onto the same model.
Tools are what turn an agent from something that writes text into something that does work. They cover calls to your own APIs, connections to MCP servers, actions that run inside your product, searching the web, handing a conversation to a human or to another agent, and changing tickets.
An agent without tools can still be useful — it reads and writes — but everything it does stays inside the conversation.
What starts a run. An agent can run on a schedule, on an event in your workspace, on an event in a connected tool, or when your own code calls it. A workflow can also hand a live customer conversation to an agent.
This is the sharpest difference from Kai, which only ever runs because a customer wrote something.
A new agent is a draft and does nothing at all until you publish it — it can't run a trigger, respond in a conversation, or even be tested. The button in the top right of the agent page publishes it, and turns into Unpublish once it's live.
Unpublishing puts the agent back to draft. That's what to reach for when an agent is behaving badly: runs stop immediately and the whole configuration stays intact, so you can fix it and publish again.
Separately from publishing, you decide who on your team may see and use an agent: everyone, only you, or specific people. Admins always have access regardless of that setting, and only the agent's creator or an admin can change it.
You don't have to assemble all this from nothing. Gleap ships templates grouped by area — support, automation, analytics, engineering, growth, sales and onboarding — covering jobs like watching SLAs, tagging tickets as they arrive, spotting churn risk, routing CSAT feedback, enriching leads and drafting a weekly performance report.
A template gives you working instructions and a sensible tool selection to edit, which is a better starting point than an empty box. You can also describe what you want in plain language and have the configuration drafted for you, or start from scratch.
