Flowgrid Agent
How answers stay accurate
The checks every agent answer passes before you see it, and why they live in code rather than in the agent's instructions.
How it works
The Flowgrid Agent answers from your website's data: it runs queries, reads recordings and builds reports. Before an answer reaches you, Flowgrid checks it. The checks are code, not instructions to the model, so they apply to every answer, in chat, in reports and in scheduled emails, and they can't be talked out of.
Under each answer, Checked N things opens the list of what the agent looked at to get there.
Why code and not the prompt
Building the agent, the fixes that held were all checks in code. The ones that lived only in the agent's instructions eventually came back.Every number comes from your data
A figure in an answer has to trace back to a query the agent ran on your website's data in that conversation.
If a number can't be traced, it isn't shown. This covers metric cards, charts and figures written in the answer itself, so an example from the agent's own instructions can never appear as one of your results.
No made-up targets or thresholds
Advice that hangs on a number you never gave Flowgrid is removed before you see it.
A model will happily suggest a benchmark that sounds researched. Flowgrid keeps the numbers from your question and the goals you set, and drops the rest, in chat answers and in every report.
“Wait until weekly visits exceed 50 before diagnosing page-level leaks.” is the kind of sentence that gets removed.
Links that exist
The agent can only link to reports and dashboard pages that are really in your Flowgrid account.
Every link is checked before it is shown: the address, and that the report it points to was actually made. A link that fails the check is removed. When the agent points you somewhere in the dashboard, it can only name a page that exists.
Recordings it has actually watched
When an answer shows session recordings, the agent has opened them and tells you what happened inside.
If the agent shows recordings without reading any, it is sent back to read them. It also can't claim more than it found: ask for the top 10 and get 3, and the answer says 3.
"Not set up" means not set up
Flowgrid only tells you something isn't installed after checking the last 90 days.
A quiet week is not a broken setup. A zero in the period you're looking at is never enough on its own to say tracking, replay or an integration is missing.
Nothing changes without you
The agent proposes. Your click makes the change.
A new board, a feature flag, a CRM update or a code fix arrives as a card for you to approve. Code changes arrive as a pull request in your repository, and nothing merges until someone on your team approves it.
When it can't answer
Some questions can't be answered from the data you have, for example about an event you don't track yet. Then the agent says so and tells you what it checked, instead of guessing. If a long investigation runs out of time, you get what it found so far, never an empty reply.
To track something new, see Custom Events.