4 min read

Release 4.2: Autopilot agents

Autopilot agents now write your recurring reports on a schedule, from the sources you choose, and wait for a person to approve them. Library search is also three times faster.

Tomás Ferreira and Hector Salinas

Product
On this page

Every team has writing that comes back on a schedule. The Monday pipeline summary. The Friday support brief. The monthly board update that takes a full day because the numbers live in four places. Release 4.2 hands that work to Autopilot agents, and keeps a person in charge of what gets sent.

This post covers what an agent is, how to set one up, what we changed in Library to make agents fast enough, and what is still on the list.

What an agent does

An agent is a saved piece of recurring writing with four parts: a schedule, a set of sources, instructions, and a destination. When the schedule fires, the agent reads its sources, drafts the document in your brand voice, checks every claim against what it read, and delivers the draft where your team already works.

A typical support agent looks like this:

  • Schedule: Fridays at 4pm, in the team’s time zone
  • Sources: this week’s Zendesk tickets tagged billing or export, plus the open bugs in Linear
  • Instructions: group tickets by root cause, name the three biggest, link examples, keep it under 400 words
  • Destination: a draft in #support-leads, with the full report in Library

Agents are built on the same cited answer engine as Research, so each sentence in the draft links back to the ticket, row or document it came from. If a claim has no source, the agent flags it instead of smoothing it over.

Review before anything is sent

We spent more time on the approval step than on the writing. In the beta, the fastest way to lose trust in an agent was a report that went out with one wrong number in it.

So every agent starts in review mode. The draft lands with a short checklist: which sources were read, which claims were checked, and anything the agent was unsure about. One person approves, edits or sends it back with a note. The note becomes part of the agent’s instructions for next time.

After a few clean runs, an owner can switch an agent to send on schedule. Even then, the agent holds the report if a source fails to load or if a figure moves more than a set threshold since the last run. A report that says “I couldn’t reach the billing export this week” is better than one that quietly leaves it out.

A faster Library

Agents read a lot. A weekly sales brief can touch two hundred call notes, so the old Library index was the slowest part of every run. For 4.2 we rebuilt it.

Release 4.1 Release 4.2
Median search, 50,000 documents 1.9 s 0.6 s
Time to index a new 100-page PDF 48 s 14 s
Agent run, weekly support brief 3 min 40 s 1 min 10 s

The new index splits long documents by section rather than by a fixed number of words, which also improved answers on contracts and research reports. Daniel wrote more about why in Long documents, short answers.

You don’t need to do anything to get the faster index. Existing workspaces were migrated over the past two weeks.

Setting up your first agent

Open Autopilot in the sidebar and choose New agent. You can start from a blank agent or from one of eleven templates, including the weekly support brief, the account research digest and the monthly metrics update.

  1. Pick the sources. Connected apps appear first; you can also point an agent at a Library folder.
  2. Write the instructions the way you’d brief a colleague. Short and specific beats long and general.
  3. Set the schedule and the destination.
  4. Run it once by hand and review the draft before you turn the schedule on.

The Autopilot product page walks through the templates in more detail.

Availability and limits

Agents are available today on the Team and Enterprise plans. Team workspaces can run up to 20 agents; Enterprise has no fixed limit. Each run counts toward your monthly usage at the same rate as a Research answer.

Admins can restrict who creates agents, which sources they can read and which channels they can post to. Every run is logged with its inputs and output, and the log is included in the audit export.

What’s next

Two things are already in progress. Agents that hand off to each other, so a research agent can feed a drafting agent. And agents that react to an event, such as a new deal closing, instead of a clock. Both will arrive in the changelog first.

If you build an agent you’re proud of, tell us. The best ones in the beta became the templates.

autopilotlibraryrelease
Portrait of Tomás Ferreira

Written by

Tomás Ferreira

Staff Engineer, Autopilot

Builds the agents that write reports while everyone else is in meetings.

Portrait of Hector Salinas

With

Hector Salinas

Co-founder and CTO

Built search infrastructure at two companies before deciding answers need sources.

All posts