Skip to main content

Agents

Automate end-to-end workflows that query your Twins, process the results, and take actions in your connected apps — on demand or on a schedule.

What's an Agent?​

An Agent is an end-to-end workflow you configure once and run whenever you need it. Where a chat is an ad-hoc back-and-forth, an Agent is a saved workflow: define what you want done once, point it at the right Twin, give it inputs, and run it on demand or on a schedule.

Agents are highly customizable:

  • They can query your own Twin or any Twin you have access to — including Expert Twins.
  • They accept variables so the same Agent can work on different inputs each run.
  • They can take actions in your connected apps through MCP tools — write to a Google Doc, send an email, post to Slack.
  • They can run on schedules (daily, weekly) or on demand.
  • They can be shared with specific people or your whole org.

Workflow Ideas to Get Started​

Weekly brag doc — schedule an Agent to query your own Twin every Friday for your accomplishments, then have it append to a running Google Doc. Pull from it at review time.

Document summarizer — an Agent with a File variable. Run it any time you need a structured summary of a doc, deck, or spreadsheet.

Meeting prep brief — a scheduled morning Agent that pulls your day's calendar and produces a one-page brief on attendees, prior decisions, and open threads.

Two tabs on the Agents page:

  • Gallery — templates plus Agents shared with you or your org.
  • My Agents — Agents you created, plus ones you've copied from the Gallery.

Click any card to open the Agent. Click Create Agent in the top-right to build one from scratch.

Creating an Agent​

Every Agent is built from the same set of pieces. You can fill them in any order before saving.

Description​

A short summary of what the Agent does. Shows up on the Agent card and helps anyone you share it with understand the purpose at a glance.

Twin to ask​

The Twin the Agent queries when it runs. Defaults to your own Twin, but you can switch to any Twin or Expert Twin you have access to — useful when the Agent's work lives in someone else's domain.

note

You can only route to Twins you already have access to. If you need access to a specific Twin or Expert, ask its owner.

Prompt​

The instructions the Agent runs. Write it like a brief: what you want, how the output should be structured, what to do with edge cases. Specific over generic.

You can reference variables inline, and you can @mention connected MCP tools to have the Agent take actions in your apps.

Variables​

Variables let you predefine inputs once and reuse them across every run of the Agent. Add a variable with a name and a type (File, String, Source), set its value, then reference it inside the prompt — for example, an Agent that reads through a specific doc each run has a document variable pointing to that file, and the prompt reads Summarize the following document: document.

At runtime, the Agent pulls from these predefined values automatically. You don't get prompted each time. If the input changes — a new team roster, a new file — update the variable once and every future run uses the new value.

Actions through connected apps​

This is where Agents get powerful. If your org has MCP connectors set up — Google Docs, Gmail, Slack, Jira, calendar, and others — your Agent can take actions in those apps as part of its run.

You invoke a connected tool by @mentioning the server and the specific tool inside the prompt. For example, an Agent can read through your week's work, write a summary to a Google Doc, and email you the link — all in one run.

Available tools depend on which MCP connectors your org has enabled. See the Apps page, or ask your admin if something you want isn't connected.

Schedules​

Schedules run the Agent on a recurring basis without you having to do anything.

Configure:

  • Prompt Variables for the run — same inputs you'd supply manually.
  • Twin to ask for this scheduled run.
  • Send to — one or more recipients who receive the output by email.
  • Email Subject (optional).
  • Frequency — One-time, Daily, or Weekly. For weekly, pick which days.
  • Time and Time Zone.

Scheduled runs deliver their output by email to the people listed in Send to. If you want a copy yourself, add yourself to the list. If the Agent's prompt also invokes MCP actions (writing to a doc, posting to Slack), those actions happen as part of the run too.

note

Email delivery is most useful for scheduled runs. For on-demand runs, you see the output live in the UI — no email needed.

Sharing an Agent​

Agents are personal by default. To share, open the Agent and use the share controls to grant access to specific people or your whole org.

People you share an Agent with can run it and copy it. If they want to modify it, they make their own copy — your original stays as-is.

Save Drafts and Run History​

Use Save to persist changes. Discard undoes everything since your last save.

The Run History tab on any Agent shows past runs, the inputs supplied, and the outputs produced — useful for debugging a misbehaving prompt or pulling up an old result.

When To Build an Agent​

Build an Agent when:

  • You find yourself running the same workflow over and over against a Twin.
  • The work has a recurring cadence (weekly brief, monthly summary, end-of-day standup).
  • You want the output to land in your tools — a doc, an email, a Slack post — not just in chat.