You ask
Tell Solen the result you want, in plain words. It starts by reading the window you were in when you asked.
How it works
How Solen for Windows does a job.
What happens between your request and Solen’s last line, taken step by step from the released 0.26.1 beta 2.
Computer use for Windows
One request.
Six steps.
Only the job’s windows
Solen may act in that window, in windows the job opens, and in windows you agree to. For any other window it asks: “May Solen work in it? Only that window, and only for this task.” A no lasts for the whole job.
Read, act, look again
It reads the window’s text and controls, takes a picture when it needs to see, then clicks, types, scrolls, drags or presses keys, and looks again. Scrolling, dragging, keys and some clicks use your real mouse and keyboard.
Ask before risky steps
Sending to other people, paying, deleting for good, and security or access changes wait for your yes, unless your request asked for exactly that step. Before any send, payment or delete, Solen compares the draft with your request. No answer within two minutes counts as no.
Check before Done
If Solen pressed something risky, such as Publish or Send, and hasn’t looked at that window since, it won’t say Done; it says Have a look. After Publish or Send it can also ask the AI model to look again at what it set. These checks can miss mistakes, so review important results.
A plain ending
Each request ends with a plain ending: Done, Answer, Have a look, Didn’t finish, Nothing was changed, or Stopped with what was already done. Then comes Solen’s answer and a count of what it did, such as “42 actions in Edge.”
Limits and stopping
You can always
stop it.
Stop anytime
Press Ctrl + Alt + Esc or the Stop button. Solen stops and says what it had already done. If you start typing or clicking while Solen is typing, its typing stops.
Time per request
By default a request can run for up to 30 minutes; the clock pauses while Solen waits for your answer. If it reaches the limit, it says so and leaves the rest open.
Questions it asks you
When Solen needs something only you know, it asks one short question and waits up to eight minutes. It never invents your answer.
Lab notes
Numbers from
our own tests.
Short findings from our research, each with what was measured. Most come from our test bench, where the result is checked in each app’s saved data. Bench timings are not the same as time on your PC.
On our blog-publishing trap, Solen said Done falsely in 8 of 16 runs. After we added a step that has it look again at what it set after Publish, 0 of 16.
Test bench, 16 runs per version, pre-release build. That step is on in 0.26.1 beta 2.
The same step took false Done from 6 to 0 across 60 runs of our 30-job bench, and finished jobs rose from 50 to 58 of 60.
Test bench, one AI model, pre-release build.
When the answer to “send?”, “pay?” or “delete?” was no, Solen stopped every time: 12 of 12 runs.
Three test-bench jobs, two AI models, two runs each, pre-release build of the new core.
A job with four parts (add two action items from an email as tasks, mark one done, reply to the sender, send) finished in 4 of 4 runs.
Test bench, two AI models, two runs each, pre-release build. One model ended with Have a look rather than Done even though the work was right.
Median time per finished job on our 30-job bench: 35 seconds for the previous design, 15 seconds for the new core on the same model, 12 seconds on a smaller model.
Test bench, pre-release build. Bench seconds are not seconds on your desk.
On real-desktop runs of the builds that became 0.26.1 beta 2, 10 of 16 genuine misses were in how the job ended, not in the clicking and typing. The 4 misses in clicking and typing were all in drawing apps.
43 labelled runs on a real Windows desktop, one run per job, one AI model through a test runner.
Very long jobs aren’t there yet: in a research test at three times our long-job length, none of 10 runs finished, across four test setups.
Test bench, one try each, November 12, 2026. It used a stand-in for Solen’s AI connection with tighter limits than the released app, and some runs were stopped at a cost limit.
How AI models compare on real tasks is on the Benchmarks page. For the basics, read what computer use is and ways to automate tasks on Windows.
Access