The legwork is done by the time you sit down.

You know this Monday.

It is 09:40. A customer email says that checkout "does not work". Support has three tickets that can be the same problem. Someone in Slack asks what changed after last week's release. You open a recording and watch a stranger fill in a form for four minutes.

By lunch, you have watched six sessions, filed one vague ticket and answered nobody. That was not the job. The job was to decide what to build next.

Autopilot gives that legwork to your agent. Sequential counts five kinds of problem in the events your app sends: errors, failed requests, dead clicks, rage clicks and dead ends. A JavaScript error or a server failure opens a problem from the first session. Other problems open when at least 3 sessions from at least 2 people hit them in the last 7 days. Each problem comes with an exact count and the sessions behind it, ready for your agent.

The same Monday, with Autopilot.

This example is illustrative. Northlight Outfitters is a fictional shop, and its numbers are made up.

When you sit down, your agent has done its 08:00 run. It left this:

• A ticket in Jira. WEB-231: People get stuck at the shipping address. 38 people. The evidence shows that the shop rejected a postcode with a space. Three sessions are attached as evidence.
• A pull request in GitHub. "Accept spaces in postcodes". Your agent opened it with its own GitHub access.
• A message in #web-team. It asks a person to review the pull request.

Your part is one review. You merge the pull request. Your deploy records release 2026.10.09 against the problem.

A week later, Sequential has counted the shipping page before and after release 2026.10.09: 41% of sessions before, 12% after. Your agent closes WEB-231 and posts the numbers. This is a measurement, not a cause. You decide what it means.

Two more mornings at Northlight

• A new error after a release. Since release 2026.10.11, a new error on search hits 21 people. Before that release, it hit nobody. Your agent opens a revert, links it to the problem and posts three sessions in #search. Your part: one approval.
• A question that the data cannot answer. Your agent's prompt asks: "Which plan do people who leave at checkout use?" The agent checks which properties Sequential has recorded. It finds that plan is not there, and it opens a pull request to send it. Your part: you asked the question.

Set it up once. Three steps.

1. 01

   Send your sessions.

   Add our script, or point a supported analytics SDK that is already in your app at Sequential. You do not re-install. Mobile apps: we test iOS with a real app on the simulator, and we did not test Android with a real app yet.

2. 02

   Connect your agent.

   An owner or admin creates a token on the Agents page. Then they add it to Claude Code, Cursor, Codex or Claude Desktop.

3. 03

   Give it a morning.

   Add one prompt to your agent's own scheduler. Sequential does not run your agent.

Every weekday at 08:00, work through what's new in Sequential.

For engineers: the docs list every tool that your agent can call.

Who does what

Autopilot, not autonomy.

• Sequential finds and measures. It counts problems and the people they hit. It records your releases and measures the same page before and after. It also ranks sessions that are worth a look, but that ranking never opens a problem.
• Your agent acts in your tools. It files tickets in Jira or Linear, opens pull requests in GitHub and posts in Slack. It uses its own connections and only the access that you gave it. Sequential never connects to those tools.
• You decide. You review, approve and merge. A draft fix goes nowhere until an owner or admin approves it. Only a person rules on a problem in Sequential.

An approved draft is a prompt for a coding agent. The agent opens the pull request, and a person reviews it. Sequential does not write or merge code.

Good to know

• You run the agent. Sequential has no scheduler. You need an agent that runs on a schedule, with its own access to your tools.
• Retention. The hosted preview keeps 7 days of data by default.
• Measurements need traffic. A quiet page can take days. If too few sessions reach the page, the result is "Not measured". With 7 days of data kept, this can happen within a week.
• What changed, never why. Each result has ranges, and notes on other changes that Sequential can see, such as another release. It cannot see all of them.
• Only a release that names the problem starts a measurement. A merge, a fix link or a ticket does not. Your deploy can record releases with its own token.
• Writes are opt-in. The recommended token only reads. An owner or admin must add each write: record a release, link a fix, ask for a draft, or propose a change to your questions.
• Daily limits. Judging and Ask have a daily limit for each project. When the limit is used up, they stop until midnight UTC. Findings and measurements continue.

Questions