One research run, unrolled stage by stage: what your agent prepares, what your user experiences inside your product, and what comes back as evidence you can check. The running example is a question most SaaS teams know too well — people start a trial, poke around, and quietly leave.
In 46 seconds: design the study, meet a user inside the product, connect their answer to the source moment, compare related interviews, and hand grounded work back to an agent.
Seven stages take an open question to work your tracker can trust. None of them is a black box — every stage leaves an artifact you or your agent can open, check, and challenge.
You → your agent
It starts with the question your analytics cannot answer. Signups look fine, activation does not, and cancellations explain nothing. You hand the question to your agent the way you hand it any other task.
The brief, as you would type it
Trials keep stalling before anyone creates a report. Find out why — with evidence, not guesses.
MCP · CLI · Dashboard
Your agent drafts a study: the goal, how to meet users — silent observation, a task to try, or an open conversation — and the debrief questions it plans to ask. You review the plan and the consent copy before anything goes live.
Study plan — ready for your review
Inside your product · Widget
There is no call to book and no panel to rent. When a trial user is in your product anyway, the interviewer meets them there — with consent — and watches the real workflow: voice, navigation, and, where the browser supports it, the screen, in one record. When the user gets stuck, nothing pops up to rescue them. Stuckness is evidence.
Captured session — minute one of six
While context is fresh
Observation first, questions second. The interviewer asks why about the exact moment it just watched — while the user can still explain what they expected. What an exit survey asks three weeks too late, the debrief asks forty seconds later.
Planned debrief
Interviewer
You clicked New report three times, then left the page. What did you expect to happen?
Participant
I thought the report would show up in that list. Nothing changed, so I assumed the trial did not include it.
Source-linked
Everything the run learned comes back attached to its source — the participant's words, the observed behavior, or both — with the page path, the playback moment, and a confidence level. Participant words stay marked as participant words; interpretation stays marked as interpretation. If a card reads wrong, the recording is one click away.
Evidence card
Nothing changed, so I assumed the trial did not include it.
/reports · playback 00:19–00:52 · interview #14
Backlog → ready → tracker
Related evidence from other interviews groups into one draft Work item. It stays a draft until someone with project context — you, or a project-aware agent — verifies the sources and moves it to ready. Only ready Work can be pushed to your tracker (GitHub or Linear), with its evidence attached, so the issue arrives explaining itself.
Draft Work — under review
Report creation gives no feedback, reads as a trial limit
Pushed to Linear with sources attached
Closing the loop
When a linked Linear issue completes, the evidence beneath it resolves — the loop closes with a receipt, not a memory. And future interviews keep watching for the same pattern: if users still stall at that moment, it comes back as new evidence, not as a hunch.
Recurrence watch
Evidence resolved when the Linear issue completed
New interviews checked for the same stall
None of these are exotic. They are the complaints every SaaS builder recognizes — and the loop above is shaped by them.
“Users churn without a word. The exit survey is a ghost town.”
What changes
Interviews happen mid-use, while users still care enough to talk — not after they have already left.
“Nobody books the fifteen-minute call. Not even the power users.”
What changes
There is nothing to book. The interview meets users inside the product, in the flow they are already in.
“Analytics shows exactly where they drop off. Never why.”
What changes
Observation captures the drop as it happens, and the debrief asks about that exact moment while it is fresh.
“Our roadmap is whoever shouted loudest in support this week.”
What changes
Evidence carries sources and confidence, so a quiet pattern across interviews can outweigh the loudest ticket.
“My agent can build anything. I still do not know what to tell it.”
What changes
The output is evidence-backed Work an agent can execute — the missing upstream input, not another dashboard.
Trust comes cheaper when the edges are drawn honestly. These are deliberate boundaries, not roadmap gaps.
It interviews users your product can already reach. No panels, no rented audiences — and no guaranteed responses if nobody shows up.
During observation nothing interrupts, hints, or navigates for the user. Stuck moments are preserved as evidence, then explained in the debrief.
Grouped evidence becomes draft Work, not ticket-ready orders. Nothing reaches GitHub or Linear until a human or project-aware agent marks it ready.
Every conclusion stays pinned to its source — a quote, observed behavior, or both — plus the page path and playback moment. When interpretation and recording disagree, the recording wins.
Ready to go deeper? The docs cover study design, the widget embed, methodology, and the agent surfaces.
Set it up in the dashboard, or let your agent do it over MCP and review the evidence when it lands. Either way, the next decision ships with sources.