Skip to main content
This guide covers the full workflow: adding an agent to an existing workspace and shaping how it behaves.

​
Add an agent

Three entry points:
  • From the Marketplace: open a specialization and use Add to Team (or create a solo agent from it).
  • From the sidebar: the Add Agent link under a team with no agents yet.
  • During team creation: the + Add Agent button on the Review screen.
The Add Agent modal asks for:
  1. Agent Name: for example “Schema Wrangler”.
  2. Source: Syndicate (built-ins), User Made, or Community.
  3. Specialization: the role and its baked-in skills, or “General-Purpose (no baked skills)”.
  4. Model: any model from a connected provider.
Click Add to Team and the agent appears on the canvas, idle and ready.

​
Configure it

Open the agent’s settings icon for the configuration panel. What to check, in order of impact:
1

Model

Match the model to the role: strongest models for planning-heavy roles, code models for coders, fast models for high-volume monitors. See Providers and models.
2

Permission mode

Pick from the shield selector: Ask first, Auto-accept edits, or Full auto. Start conservative; loosen as trust builds. See Permissions and safety.
3

Soul

The agent’s system prompt. For user-made agents, click Edit to add project-specific guidance. Effective Soul additions are concrete:Weak: “Write good code.”Better: “This project uses React 19 with Tailwind. Prefer existing components in src/components. Never introduce new dependencies without flagging it to Atlas first.”
4

Skills

Review the Skills (N) list so you know which slash commands the agent answers to. Skills come from the specialization; to change the set, change the specialization or build a custom package with Agent Creator.
5

Tools (MCP)

Toggle on only the MCP servers this role needs. Servers are installed globally in Settings, MCP, then enabled per agent here. See MCP servers.
6

Context

The defaults (Balanced) are right for most agents. If runs feel slow or the agent misses recent instructions, open the Context editor and check Prompt Analytics to see where the tokens go. See Agents.

​
Verify

Send the agent one small, role-typical message and confirm it responds in character, uses the right tools, and respects its permission mode. Then put it to work.

​
Remove an agent

Configuration panel, Danger Zone, Delete Agent (two-step confirmation): “Remove this agent and all its config.” Queued tasks assigned to it are removed from the board.