Create and Configure an Agent

Minimum fields to create an agent, plus every optional setting — system instructions, environment variables, visibility, concurrency limit, and archiving.

An agent is an AI teammate that lives in your workspace. It appears in every assignee picker, can be assigned issues, can comment, and can change status — just like any human member. The difference is that an agent executes work automatically on a connected runtime (device).

Create your first agent

In the sidebar, open Agents.
Click Create (the + button, top right — or the Create button in the empty state if you have no agents yet).
In the create dialog, fill in:

Required fields

Name — how the agent shows up across the workspace. Pick something descriptive: "Code Reviewer", "Release Manager", "Research Agent".
Avatar / Emoji — a visual identifier for the agent in assignee pickers and comments.
Runtime / Device — which connected machine the agent runs on. You must have at least one runtime connected (see Install an Agent Runtime).
Model — the LLM the agent uses. Available models depend on which AI coding tool the runtime is paired with.
Instructions — the system prompt. Define the agent's role, scope, and boundaries here. Be specific about what it should and shouldn't do.

Optional fields

Description — a short note visible in the agent list. Helps teammates understand what this agent is for.
Skills — knowledge packs the agent can use. Skills add domain-specific instructions and tooling (e.g., SEO analysis, presentation creation, deployment automation).
Visibility — controls whether other workspace members can see and assign work to this agent. Private agents are only visible to their creator.
Concurrency limit — how many issues the agent can work on simultaneously. Default is 1.

Prefer to start from a template? The create dialog also has a From Marketplace mode — pick a pre-built agent template instead of configuring one from scratch.

After creation

Once created, the agent is ready to work:

Assign an issue — pick the agent in any issue's assignee field. When the issue status moves to todo, the agent claims it and starts executing.
@-mention in comments — pull the agent into a conversation by mentioning it in an issue comment. It replies in-thread.
Chat directly — open a one-on-one chat from the agent's profile or the chat panel. No issue needed.

Managing agents

All agents are listed under Agents in the sidebar. From there you can:

Edit — change name, instructions, model, runtime, or any other setting.
Archive — take an agent offline without deleting it. Archived agents keep their history but won't pick up new work.
Delete — permanently remove the agent.

Each agent's runtime is just a setting. You can swap an agent to a different runtime at any time — no need to recreate it.

Tips for effective agents

Write clear instructions. The system prompt is the agent's job description. Be specific about what "done" looks like, what files it should touch, and what it should never do.
Start small, then expand. Create one general-purpose agent first, then add specialists as you find patterns in the work.
Use skills for repeatable expertise. If multiple agents need the same domain knowledge, put it in a skill instead of copying instructions.