Build log
Build log.
Notes from building an agentic browser in the open — architecture decisions, security work, and the experiments that did not work.
Editorial policy
What gets published here, and what does not.
- Check mark: Publish the refutations. When a designed approach is tested and fails, that post gets written. One phase of this project records a measured refutation of its own specification; that is the standard.
- Check mark: No announcement without an artifact. A post claiming a result links to the thing that produced it.
- Check mark: Cadence over polish. A short honest note every few weeks beats a quarterly essay.
- Check mark: No competitor dunking. Published incidents in the category are fair to analyse, technically and without gloating — they are how the whole field learns what an agentic browser must not do.
Published
Out now.
The screenshot that captured the wrong screen
Why a browser whose tabs are isolated WebContentsViews cannot screenshot itself, the OS-level workaround that captured the operator's own desktop twice and was deleted, and the desktopCapturer path that works. Ends where it actually ended: the capture is solved, the agent recording is not. Read it.
In the pipeline
Drawn from work already in the repository.
Six posts that exist as material — each has a real conclusion, not a teaser. They land here as they are written.
- Why the model is not the security control. The failure shape behind the category's public incidents, and what a deterministic kernel that runs before the model changes about it.
- We tested our own sandbox design and it failed. A phase specified an isolated-world approach for code execution. A cheap offline experiment refuted it. Why that got recorded as a result rather than quietly redesigned around.
- A kill switch that fails closed, and how we proved it. Killing a live tunnel endpoint against the built application and confirming that a proven-reachable clear path records nothing. Includes the residual leak we could not close and documented instead.
- Deleting a native database. Replacing
better-sqlite3with Node's built-in SQLite removed a rebuild script, three CI steps, and a skip-guard that had been letting 63 tests sit out entire runs. - What a browser agent must never do on its own. Enabling a sensitive category, widening a wallet mandate, and permission widening in general — the list, and why every one of them belongs to the user rather than the agent.
- Reading everyone else's complaints. What the published user studies of five rival agentic browsers say, what we changed because of them, and the ones we have not answered yet.
Follow along
No email list yet.
An email option makes this site a data controller, and the privacy policy has to be final before that field exists. Until then: watch the repository, or check back — the cadence is a note every few weeks.