Find out in five minutes whether the drop is real or a renamed event

The twelve per cent drop that might be a tracking bug

The companion starts where an engineer would.

Raw events for the signup funnel across both weeks, event names compared, and there it is: `signup_completed` stopped appearing on the 14th while a similarly named event started the same day. Five minutes for what used to take an hour, and the query comes with it so you can disagree with the method rather than the conclusion.

When the drop is real, the same pass continues: which acquisition sources moved, which persons in the affected segment did what instead. That is the throwaway analysis that precedes a dashboard, and it is exactly the work worth handing off.

How does a companion write HogQL it can defend?

By reading first.

Saved insights show how your team already models funnels, trends, retention and paths. The event list shows what is actually being sent rather than what the tracking plan says should be. A query built from those two uses your definitions.

Then results come back with the HogQL that produced them, so a result you distrust is a query you can read. That habit is what makes this research something you can put in front of a team.

Should an agent be allowed to flip a feature flag?

It can create and update flags, and a flag change takes effect for real users the moment it lands. So flag writes are staged and wait for your explicit approval, with the flag key and the new state written out plainly before you confirm.

Inside that gate it saves real time: “create the flag for the new onboarding, off for everyone, with a note explaining what it gates” is one instruction, and rollout changes are two words each.

What happens in the PostHog interface itself

Queries, events, persons, insights and flags are native and repeatable, which is why the drop investigation above is a five-minute pass you can schedule.

Session replays, dashboard editing, surveys and experiment configuration live in the interface, so a companion opens them in the PostHog you are signed into and works through them as you would, with you watching. Either half lands the same way: structured results written where your team reads them, with the query or the source attached to every row, the ordinary data extraction loop.

Frequently asked questions