Agents

An agent is a named assistant with a model, personality, and tools. New accounts start with one primary assistant. Switch to an ensemble when you want separate specialists; see Single assistant and ensembles.

What an agent has

  • Name and emoji — how you identify them in the sidebar and in cross-agent delegations.
  • Role — the persona and workflow rules they follow (see Roles). Determines which built-in skills they’re allowed to use.
  • Model — one model from one enabled provider. Can be changed any time.
  • Description — a short routing hint; include specific words you expect to use in requests (for example, UI, CSS, API, or database). The Coordinator uses those terms when deciding who to delegate to.
  • Personality — optional; how the agent talks. “Warm and playful, loves a good pun”, “dry and blunt, zero fluff”, “explains like a patient teacher”. It shapes tone and style in every reply (voice included) without changing what the agent can do — capabilities stay with the role and tools.
  • Tools — the specific skill tools this agent has unlocked. Roles set sensible defaults; you can adjust per-agent.

Creating an agent

In Single assistant mode, you can create your first assistant. To add another, switch to Settings → Agents → Agent setup → Agent ensemble first.

Two ways:

  • Sidebar → Agents → + New Agent.
  • Or type /new-agent in the chat input — same modal, faster if you’re already typing.

Pick a role first — that pre-fills the tools and system prompt. Save. The agent appears in the list immediately.

Talking directly vs. through the Coordinator

  • Click an agent in the sidebar list to open a direct chat. This conversation has a persistent session — it remembers context across messages.
  • Talk to the Coordinator instead and it will delegate — each delegation creates an ephemeral session for the chosen specialist, just for that task. The Coordinator collects the result and replies to you.

That distinction matters: when you ask the Coordinator to “have the researcher look up X”, the researcher gets a clean fresh context every time. To accumulate state with a specialist, talk to it directly.

Editing an agent later

In the agent list, hover the agent and click the edit icon (or right-click → Edit). You can change name, emoji, model, description, personality, and tool selection without losing chat history. Personality edits apply from your next message — no restart or new conversation needed. Switching role (via role assignment) swaps the workflow rules without rebuilding the agent.

Scheduled tasks owned by an agent always use that agent’s current model — the model isn’t snapshotted at schedule time. So if you swap an agent from gpt-5.5 to claude-sonnet-4.6 today, every future fire of every scheduled task that agent owns runs on the new model. Same for watch on-fire actions.

Expand the Advanced dropdown in the edit panel for the less-touched knobs:

  • Context window — how many prior tokens of conversation to include each turn. Lower it if responses are slow or you’re getting close to the model’s limit; raise it if the agent forgets things mid-task.
  • Max output tokens — caps the length of any single response. Useful if an agent rambles or to keep cost predictable on long-output models.

These are per-agent, so you can give a researcher generous limits and a quick-reply assistant tighter ones.

Deleting an agent

Edit panel → Delete. Sessions, history, and any per-agent state under users/{userId}/agents/{agentId}/ are removed.

Multiple agents on the same model

Totally fine, and common. Two coders on the same OpenAI model with different system prompts (one front-end, one infra) is a perfectly good setup.

This site uses Just the Docs, a documentation theme for Jekyll.