Self-driving 101
Self-driving is PostHog using your product data as context to proactively improve your product. Here's how it works:
You're in control to review and merge everything. Nothing ships without you.
The loop
The whole system is Fig. 1. A signal is something in your product data that's worth knowing, with its evidence attached. Related signals group into a report for you to review. An agent writes the fix as a pull request, and PostHog measures whether it worked.
the outcome becomes new signals
Here are some examples of signals:
- Your code throws a new error in production, captured by Error Tracking.
- Signups, conversions, or retention dip below their usual range.
- Users hesitate or get lost in session replays, which Replay Vision watches so you don't have to.
- A ticket or issue lands in a connected tool like Zendesk, Linear, or Sentry.
The reports created from signals land in your inbox. Each one has an anatomy like this:
| Content | Description |
|---|---|
| Grouped signals | The related findings, from one or more sources, with the evidence behind them. |
| Priority | How urgent it is, so the most important work rises to the top. |
| State | Whether it's actionable, so an agent can open a pull request, or needs input before anything can be worked on. |
| Charts | Where a finding is about something that moved, the report can carry the chart that shows it, so you can see the shape of the problem without rebuilding the query yourself. |
High-confidence reports arrive with the pull request already open, like this one from PostHog. You just review and merge. Lower-confidence ones wait in your inbox for you to decide first.
Once the PR merges, the change generates new signals, so your product keeps improving itself over time.
What's acted on
Each time a scout runs, it ends one of two ways: it writes a report, or it writes nothing.
A scout only writes a report when it spots a change that needs your attention:
- Your AI bill jumped.
- A core activation metric dipped.
- A new error showed up in production.
It stays quiet when everything looks healthy, including when a trend indicates growth or noise.
Quiet is the common case. A scout that reports every day is one you'll mute before you get real value out of it.
Apply a use case example
To help you self-drive, we've included a few use cases in this pocket guide. Each use case is a real custom scout you can add to your own product. Every guide covers:
- The scout's job, and the anatomy of the inbox report it files
- The pull request opened based on the reports
- The scout itself – the actual file the agent follows
- A one-click way to add it to your own product