Creating and Configuring Agents

An agent is an AI identity: a personality, skills, tools, and guardrails. Your personal agent — pinned at the top of the sidebar — is created with your account and is the default place to chat and delegate. Workspace agents are the ones you create for specific jobs.

The identity part matters more than it first appears. A one-off chat starts from zero every time; an agent accumulates. Its personality shapes every conversation, its memory carries what it learned last week into next week’s work, and its assignments give it a standing job rather than a stream of ad-hoc requests. Over time an agent stops feeling like a chat window and starts feeling like a colleague with a beat: the one who watches competitors, the one who runs the Monday pipeline report, the one who triages inbound tickets.

Creating an agent

Choose New agent on the Agents page (or New agent in the sidebar), give it a Name, and describe it under Describe it: the work, who it’s for, how often, and how it should come across. That’s the whole form — everything else is configured on the agent’s page.

The description becomes the agent’s first message. It writes its own personality from what you said and shows you the text, then proposes an assignment for any work that recurs and schedules it once you agree. To set the agent up by hand instead, choose Write it yourself and enter the personality directly.

Names deserve a moment of thought, since they are how the agent appears across sessions and the Inbox, and how you @-mention it in a conversation. Short, role-shaped names wear well:

• Scout — watches the competitive landscape and maintains a tracker table.
• Pipeline — compiles the weekly sales report every Monday.
• Triage — fields support tickets delivered by a webhook as they arrive.

The personality is the agent’s standing character: how it writes, what it cares about, how it approaches work. A useful one reads like a briefing you’d give a new hire rather than configuration. For example:

You are Scout, a competitive-intelligence analyst. You are skeptical of vendor announcements until you can confirm them from a second source, and you always link the source. You write in short, factual sentences and lead with what changed since last time. When you are uncertain, say so plainly instead of hedging. Prefer a table over prose for anything with more than three items.

Start with your personal agent

You don’t need a roster on day one. Chat and delegate through your personal agent first; when you notice a job recurring — the same report, the same research, the same triage — that’s the moment to create a dedicated workspace agent for it, or to turn the conversation itself into an assignment with Save as Assignment.

The agent page

The main column of an agent’s page starts with a composer: write to the agent and the conversation opens in its own session. Beneath it is everything the agent has done that you can see, newest first — its assignment runs, your conversations with it, and sessions others have shared — each with how it went and the work it produced. The right side is a stack of cards for its standing work and its configuration.

Personality sits in the header, always in view: a preview of how this agent thinks, communicates, and approaches tasks, applied to every conversation. Click it to edit — the editor includes AI refinement, and an agent allowed to adjust its own personality will also refine it as it learns. Whether it may do that is a self-improvement guardrail, so you decide which agents earn that latitude.

The cards on the right each show a one-line summary and open to the detail:

• Assignments — the agent’s standing work: how many it has and when the next one runs. Open it to run, pause, or add one. See Assignments.
• Abilities — the skills, integrations, and private credentials this agent uses. See Skills and Integrations.
• Memory — notes the agent keeps for itself as it works. See Memory and storage.
• Storage — the files the agent keeps between sessions.
• Guardrails — limits on what this agent can reach or change. See Guardrails.
• API — on workspace agents: give the agent an API so other tools and agents can call it. See Agent API.
• Messaging — how to reach this agent outside the app, by its own email address, a text, or a Slack DM, and how it may reach you. See Inbox.

Together they answer the questions you would ask about any coworker: what is this agent responsible for (Assignments), what can it do (Abilities), what has it learned and what is it working on (Memory and Storage), and what is it not allowed to do (Guardrails).

A status line under the name keeps you oriented — whether the agent is running or idle, its current session, and when its next scheduled run is due. The Automated runs switch beside the Assignments heading pauses every automated trigger while still allowing manual runs — useful while you rewrite an assignment’s instructions and don’t want the old version firing on schedule mid-edit.

Sharing an agent

Workspace agents start private to you, which makes them a safe place to iterate: tune the personality, run a few sessions, adjust the guardrails, all before anyone else depends on the results. A private agent runs as you; workspace admins can see its name, connections, and last run, but never its sessions, memory, or instructions.

The Share button in the agent’s header lists who has access. Add specific people at Can view, Can use, or Can edit. Opening the agent to the whole workspace is a publishing step, and you choose between two shapes:

• Agent — one agent everyone talks to. It takes its personality, skills, assignments, and memory to the team and works on company accounts from then on; your own accounts stay yours. Once published, you set the level everyone in the workspace gets.
• Template — anyone can adopt their own copy, which works as them, and you keep yours. You can give every member a copy now and as they join. Publishing as a template can’t be undone.

Granting the whole workspace use while keeping edit to one or two people is a common shape: the team gets the agent, and its personality, assignments, and guardrails stay coherent.

Star the agents you use most to keep them at hand.

Patterns that work

• One agent per job — an agent with a single clear beat (competitors, pipeline, tickets) develops sharper memory and a more useful personality than a generalist asked to do everything.
• Personal agent as front door — delegate everything there first, and spin out a workspace agent when a duty becomes recurring.
• Shared team agent — one well-tuned agent shared at the use tier, with edit access kept to its owners.