How to run interviews, analyse what people say, and turn it into decisions. Written by the team behind User Evaluation.
Two quieter changes this month. Deleting a file, study, report, clip or reel now removes its stored files too, not just the entry you saw. And adding a whole project as context for Eva is a searchable, multi-select dialog in the composer.
AI agents that use your website and real people who use your product answer different questions. Here is a practical guide to which one to reach for, what each is bad at, and how to combine them so agents find the problems and people decide which ones matter.
AI agents are good at finding where a website breaks. Only people can tell you whether it matters. One button on a Swarm run now drafts a short study that asks 5 real people about every finding, ready for you to check and launch.
Five agents hitting the same bug used to mean five findings worded five ways. Swarm now triages a run's findings into one clean set with a short summary, shows the real screenshot from every session that hit each one, and puts each issue on the flow map next to the screen where it happened.
A Swarm run is no longer a box you check back on later. The run page now shows agents working in real time, the replay follows the newest step as it happens, and any session plays back in seconds once it is done.
Swarm agents now look at a screenshot of the page every turn and read the errors a developer would see: console errors, failed requests and slow requests. So a run reports broken layouts and technical faults, not only confusing steps.
Websites treat scripts differently from people, so a test that looks like a script tests the wrong thing. Swarm agents now use a real, visible browser, move the mouse and type the way people do, and wait out slow pages instead of giving up.
Ask Eva for a Swarm run and the run card appears straight away, with 5 agents, a desktop browser and one visit each. The only thing she needs from you is the website. Everything else is one click away in tabbed run settings.
A captcha used to end an AI agent's session. Now the agent raises a hand, you open its live browser, solve the check yourself, and hand the page back. The agent carries on from where it stopped.
