Build an agent

An agent is a worker with its own instructions, skills, connections and cloud computer. When no agent in your workspace fits a job, Kav has one built for it. This page covers what to tell Kav and how to give a new agent real expertise.

Problems this solves

Describe the job

Tell Kav what the agent should do and what a good result looks like. The builder fills in the rest.

Build me an agent that answers billing questions in our help desk, so fewer billing
tickets reach the team. A good reply cites our pricing page and never promises a refund
without my approval. Work from our pricing page and docs/refunds.md.

Kav proposes a name, and once you agree, files a "Build" task you can watch on the calendar. When the agent should run, on a schedule or when something happens, is set on a workflow. See Write a workflow.

What the builder does

1. Names the goal the agent works toward.
2. Researches the field: the core craft, the expertise a strong practitioner also relies on, and the moves specific to your goal. It works from recognised sources and skips listicles and vendor marketing.
3. Collects 8 to 10 real, proven examples when the output is judged on taste, such as a video, an ad, a design or a cold email.
4. Installs the expertise as skills, keeping a practitioner's published skill package whole.
5. Writes the agent's instructions.
6. Asks you, once, whether to use an account you already have for each service the job needs, or one it registers for you.
7. Tests the agent and hands it back on the board.

Point Kav at expertise worth copying

The fastest way to a strong agent is a method that a recognised practitioner already wrote down. Paste a link to a folder of skills on GitHub, and ask for an agent that follows it:

Build an engineering agent that follows Lauren Tan's pstack:
https://github.com/cursor/plugins/tree/main/pstack

The builder clones the repository, reads its skills and installs them on the agent. A skill folder with its own scripts or examples is kept whole.

Other good sources are the skills published by the makers of the tools an agent uses. For example, Cloudflare publishes skills for Workers and Wrangler at github.com/cloudflare/skills.

Start from a pack

A pack is a set of agents that come with their skills and the apps they use. The engineering pack has two agents:

Add a pack in the Choose agent packs step when you set up a workspace, or from kavela.ai/marketplace. Adding one opens its setup under Agents and starts a chat with Kav to connect the apps it needs.

Test it on one run

Before an agent works on its own, ask for one run and read the result: "Do one run of this now and show me what you'd deliver." See Refine the output.

Find and edit an agent

Open Agents in the sidebar. Each agent's page has:

To change an agent in plain words, tell Kav, for example: "Have Ship add a checklist step before every deploy." Kav files the change as a task for the builder.