Agents

Spawning Agents

Agents are the atomic unit in SenateLab. Each one is a scoped, named process with its own memory and permissions. Onboarding is a four-step wizard; after that, you can refine anything from the detail panel.

The four-step onboarding

1. Role. Pick from Manager, Engineer, Researcher, Reviewer, Ops, or Support. Roles seed the other three steps with sensible defaults, which you can override.
2. Folders. Pick which folders attached to the current map this agent can read and/or write.
3. Identity. Set the soul (personality, expertise, tone), the purpose (what it should and should not do), and the display name.
4. Guardrails. Review the defaults from the map's settings and tighten any that matter for this role — for instance, a Researcher that should never write to disk.

Agent isolation

Every agent runs in its own Codex CLI session and its own temporary worktree. The session starts with only the folders you granted it; everything else is invisible. Command execution is gated by the guardrails pipeline — a shell call goes through a classifier before it leaves the sandbox, and anything that matches a guardrail waits for your approval.

Rule of thumb: give each agent the narrowest scope that lets it succeed. Wide scopes are easy to set but hard to unset later.

Customize an agent

After launch, open the detail panel and use the Identity tab to refine the soul and purpose. The Settings tab exposes model, temperature, and optional system-prompt additions. Changes take effect on the next turn — there is no restart.

Deleting an agent

Right-click the node on the canvas or the entry in the sidebar, pick Delete, and confirm. SenateLab asks whether you want to Keep the chat history (exported as a Markdown file in the map's archive) or Remove it entirely. Memory is removed either way.