Agents

An agent is a configuration that tells Crew how to behave: which model to use, what system prompt to follow, which tools it can call, and which MCP servers to spawn. Every session runs under an agent.

Open Agent Capabilities → Agents. You can:

• View all configured agents
• Edit any agent's model, prompt, and tool list
• Create new agents
• Switch the active agent for a session
• Reset a managed agent to its defaults

To use Crew members, open Settings → Developer and enable Crew Members and Crew Mode under Feature Previews.

In Agent Capabilities → Agents, open a member's Avatar → Customize control. You can choose its ghost traits one at a time or upload a picture. On the Expressions pane, set eyes, mouth, and a sound cue for working, done, and failed states.

The member drawer shows recent activity, its auto patrol, and the schedules and webhooks that can wake it. You can star a member and filter the roster by Starred, Mine, Built-in, or From packages.

On desktop, enable Crew Companion from Apps → Library to show one member avatar across your displays.

Crew ships with a kirocrew agent that includes:

• Model — auto (uses the model configured in kiro-cli)
• System prompt — general-purpose assistant behavior
• MCP servers — kirocrew-core (spawn, learn, task, wait, messaging) and kirocrew-cron (scheduling tools)
• Tools — full tool access (file read/write, shell, web, MCP)

This is what runs when you don't specify an agent. It covers most use cases.

From the Agents panel, click Create. Set:

• Name — identifier used to reference the agent
• Model — which LLM to use (or auto for the default)
• Prompt — the system prompt that shapes behavior
• Tools — which tools the agent can call

For file-based configuration, create a JSON file:

{
  "name": "reviewer",
  "description": "Independent code reviewer",
  "model": "auto",
  "prompt": "You are a code reviewer. Focus on correctness, security, and clarity.",
  "tools": ["fs_read", "grep"]
}

Place it under ~/.kiro/agents/ or reference it from the dashboard.

Control how the agent gets tool permission. Set globally from Settings → Security or per-session with the Autopilot toggle.

Denied commands and sensitive-path blocks always apply, even in Autopilot.

• Narrower tool access — a "reviewer" agent that can only read files, not write them
• Different prompt — an agent tuned for a specific domain (data analysis, writing, ops)
• Different model — use a specific model for certain types of work
• Restricted scope — limit what the agent can do for safety-sensitive workflows