Skip to content

Creating an agent

An agent is a named assistant that belongs to your organization. It holds instructions, a provider and model, and the apps, skills and databases it may use. Members chat with it in threads, and an Agent step in a workflow runs the same agent with a task.

Every member can create, chat with, edit and delete an agent. Agents count against your plan: Free includes 3 and Pro includes 25. See Plans.

  1. Open the Agents page and press New agent.
  2. Enter a Name and a Description.
  3. Write the Instructions, or press Start from a template.
  4. Pick a Provider, a Model and how much Thinking before each reply it does.
  5. Under Connected apps, add the connections it may act on. See Grants and approvals.
  6. Under Skills and Databases, add what it may use.
  7. Press Create agent. The agent’s page opens on its Chat tab.

The name is a slug: lowercase letters, numbers and single hyphens, up to 64 characters. Spaces become hyphens as you type. No two agents in an organization share a name.

The description is a short line, up to 1,024 characters. It shows in the agents list as the agent’s purpose. When it is empty, the list shows the first line of the instructions instead.

The instructions say what the agent is for and how it should behave. They are a Liquid template: tags fill in each time the agent’s prompt is built, so the agent always reads its current name, the date and what it may use. Type {{ in the field to pick a variable.

Variable What it holds
today The date in UTC, like Monday 2026-10-05 (UTC)
agent.name The agent’s name
agent.description The agent’s description
organization.name Your organization’s name
apps Each grant: name, connection and tools
connection On each app: the connection’s id and account
tools On each app: every granted action’s name, ref and askFirst
skills Each mounted skill: name and description
databases Each database: name and writable

Loops, conditions and filters work as they do in workflow fields. See Templates and variables. A template that doesn’t parse is refused when you save, with Liquid’s message under the field.

Start from a template replaces the instructions with one of four starting points:

  • General assistant lists every app, skill and database the agent has.
  • Customer support escalates, and explains the actions that need approval.
  • Daily briefing is built on today’s date and the apps it can read.
  • Data analyst splits read-only databases from writable ones.

Improve with AI opens the assistant on your draft. It drafts or rewrites the instructions in the field. Improve with AI is on the Pro plan.

When a template or the assistant replaces text you had, a toast offers Undo. Nothing is saved until you press Create agent or Save changes.

Beyond what you grant, every agent has the same built-in tools. The configuration panel on the agent’s page lists them as always on.

  • Workspace tools: files, a shell, TypeScript it runs in a sandbox, and share links to its files. The workspace is shared by all of the agent’s threads.
  • Browser tools: reading a public page or API, a page as Markdown, its links, scraping elements and extracting data.
  • Memory: a note for the current thread, and a shared note every thread reads and writes.

An agent acts in your apps only through its grants, and in a database only where you added it. See Chatting with an agent for memory and files.

Open the agent, choose the Editor tab, make your changes and press Save changes. A saved change applies to the agent’s open threads at their next message, and to the next run of every workflow that uses the agent.

On the Editor tab, press Delete, then Delete agent. Every thread and every file in its workspace go with it, for every member. It can’t be undone. A workflow whose Agent step picked it can’t be published until you choose another agent, and its live version fails at that step.

Agents are part of the API: list, create, read, update and delete them with an organization key. See the API reference.