Agents

Agent Config Fields

Field	Required	Description
name	yes	Display name for the agent
driver	yes	Which CLI to use (claude_sdk, copilot_sdk, codex_sdk)
model	no	Model override (driver default used if omitted)
system_prompt	no	Additional instructions appended to the agent’s prompt
mcpServers	no	MCP tool servers available to this agent (see MCP Servers)

Roles and the Agent Pool

Roles are not configured on the agent. When you create an execution, you select a root agent and an agent pool — the set of agents available for delegation.

The system assigns roles based on where an agent ends up in the delegation tree:

• root-lead — the agent you selected as root. Receives the execution prompt, coordinates the team.
• sub-lead — an agent that was delegated to and can further delegate.
• leaf — an agent at the maximum delegation depth. Executes tasks directly.

The same agent config can play different roles in different executions depending on how the lead chooses to delegate.

Example: Minimal Claude Agent

{
  "name": "claude-worker",
  "driver": "claude_sdk"
}

Example: Agent with Custom Instructions

{
  "name": "security-reviewer",
  "driver": "claude_sdk",
  "model": "sonnet",
  "system_prompt": "You are a security-focused code reviewer. Flag OWASP top 10 vulnerabilities, injection risks, and authentication bypasses. Be specific about the risk and fix."
}

Briefings

Agents receive a briefing when assigned work. Briefings include:

• Project-level context (conventions, architecture notes, constraints)
• The execution prompt and any parent context
• Wiki pages the project has marked as relevant

Briefings are composed automatically — you configure the project-level content in the project settings. Agents see the briefing as part of their system prompt when they start working.