Skip to content

Agents

An agent is a customizable AI persona that runs through EM Copilot. Each agent has its own name, emoji, short label, description, and system prompt, plus an optional set of attached skills. When you talk to an agent, you’re talking to whichever persona you picked — the agent decides tone, focus, and which skills to invoke.

Agents do not carry their own model. They inherit the workspace defaults configured in Settings → EM Copilot (provider, default model, reasoning effort). Individual automations can still override the model and effort per run.

Every workspace ships with six built-in agents. One is always the workspace default (the agent used when you start a new conversation without picking one).

AgentUse it for
🗂️ Executive AssistantTask management, drafting, lookups across workspace data. Default.
🧭 Coach1:1 prep, feedback conversations, development discussions.
🥊 Sparring PartnerAdversarial review — stress-test ideas, challenge assumptions.
📈 Career AdvisorGrowth plans, leveling, promotion readiness.
✍️ WriterDrafting messages, reviews, announcements.
🔍 ResearcherSynthesis across feedback, work logs, and notes.

Built-in agents can be edited and archived but never deleted. The workspace default cannot be archived or deleted while it’s the default.

Open Settings → Agents to manage agents.

Click New agent to open the form:

  • Name — Full name shown in pickers (e.g. “Coach”).
  • Short label — Compact label prefixed to responses (e.g. “Coach”).
  • Emoji — Visual identifier shown alongside the name everywhere the agent appears.
  • Description — One-line “when to use” hint shown in the agent picker.
  • Skills — Optional multi-select of skills the agent can invoke.
  • System prompt — Markdown editor. The persona, voice, and operating principles the agent should follow on every turn.

Name, short label, and system prompt are required.

Each row in the table has actions:

  • Edit — Open the form pre-filled with the agent’s current values.
  • Duplicate — Create a copy as [Name] (copy) with all skills attached.
  • Set as default — Make this agent the workspace default for new conversations.
  • Archive / Restore — Hide an agent without deleting it. Archived agents do not appear in pickers. The current default cannot be archived.
  • Delete — Permanently remove. Only available for non-built-in, non-default agents.
  • Show / Hide archived — Toggle the table to include archived agents.

Agents surface in three places in EM Kit.

The Copilot drawer has an agent dropdown in the draft bar at the bottom. New conversations start with the workspace default agent; pick another one before sending to talk to a different persona. The agent’s emoji and short label appear next to each response so you always know who you’re talking to.

When you create or edit an automation, the Agent, model and effort section lets you pick which agent runs the automation. The chosen agent’s persona and attached skills apply on every scheduled run, so a Coach-agent automation can lean on its 1:1 prep skill, while a Researcher-agent automation produces a different tone for the same prompt.

The agent is required. Model and reasoning effort are optional overrides — leave them empty to inherit the workspace defaults.

Open any issue and click Ask Copilot in the top-right of the detail drawer. The dropdown lists every available agent (the default is marked); pick one to insert an @-mention into a new comment, then add your question and post the comment.

The mentioned agent reads the issue context (title, description, sub-issues, prior comments) and posts a reply directly in the comment thread. You can keep the conversation going by mentioning an agent again in any follow-up comment.

Agents inherit the workspace AI defaults. To change them, open Settings → EM Copilot and adjust Default model and Reasoning effort — see EM Copilot for the full configuration walkthrough.

Per-conversation overrides are available in the Copilot drawer (session-only — they do not persist) and per-automation overrides are saved on each automation. Agent definitions themselves stay model-agnostic.