Skip to main content
An agent is a single AI worker. Agents can run inside a team or as a solo workspace.

​
Anatomy

​
Team Agents vs Solo Agents

Solo Chat mode is best for questions, planning, writing, research, and review. Solo Build mode lets the agent edit files and run commands in its folder.

​
Statuses

Agents show a live status in the sidebar, canvas, and headers:
  • Active: currently working.
  • Idle: waiting for work.
  • Waiting: needs approval or input.
  • Blocked: hit an error; open chat for details.

​
Configuration Panel

Open an agent’s settings to tune it.

​
Identity

Name, specialization badge, and model selector. Models are listed by connected provider.

​
Soul

The agent’s system prompt as markdown. Built-in specializations are read-only. User-made agents expose an editable section.
Saved Soul edits cannot be undone automatically. Copy the current text somewhere safe before large rewrites.

​
Context

Controls what the agent sees on each run:
  • History & Memory: message lookback, max characters per message, and role filter.
  • Scope & Budget: team-message lookback, reference max, and token budget.
  • Sections: Team, Connectors, Tasks, Messages, References.
  • Prompt Analytics: prompt size, estimated tokens, and per-section breakdown.

​
Skills

Skills are packaged workflows agents can run with slash commands.
  • Built-in specializations include curated skills.
  • User-made packages come from Agent Creator.
  • Community packages come from GitHub, local folders, or CLI install commands.
Type / in chat to open the slash-command menu. You can also ask in plain language; slash commands are the explicit way to trigger a workflow. Suggested workflows appear in the Marketplace for built-in specializations.

​
Tools (MCP)

MCP servers connect agents to external tools and data, such as GitHub, databases, or internal APIs. Install MCP servers once in Settings, then enable them per agent. Enable only what the role needs; fewer tools means better tool choices and a smaller blast radius. Common setup paths:
  • Catalog: one-click presets for common servers.
  • Manual server: local command, remote HTTP URL, or remote SSE URL.
  • Secrets: encrypted locally and injected at runtime.
MCP tool calls go through the same permission flow as built-in tools.

​
Danger Zone

Delete Agent removes the agent and all of its configuration. Project files are not deleted.

​
Permission Mode

Set permission mode from the shield selector: Details and recommendations live in Permissions and safety.

​
Next Steps

Add and configure agents

The workflow for adding and tuning an agent.

Agent Marketplace

Browse built-in, user-made, and community specializations.

Skills catalog

Lookup every built-in skill.

Permissions and safety

Approval flows and safety controls.