When the browser becomes the agent workbench
Zephium, BrowserOS neo, and Opera place sessions, models, approvals, and human review at different layers of the browsing experience.
Tools in this story
- Zephium — Package details and install
- BrowserOS neo — Package details and install
- Opera — Package details and install

The browser is no longer just the destination
Browser agents used to arrive from outside: a script launched a headless engine, performed a task, and returned a result. Zephium starts from a different premise. It is a complete browser whose Work canvas lets agents operate beside the person using it. Sources, intermediate pages, questions, and results remain visible instead of disappearing behind a remote automation job. The browser can also be used with AI disabled, which makes the agent layer a choice rather than the price of entry.
Zephium is genuinely new. Its public repository began in June 2026, and the upstream release history reached a stable v1.0.3 in October. That does not establish long-term reliability, but it does satisfy the editorial policy's definition of a newer tool. More important is what the documentation makes inspectable: the app uses WebKit on macOS and WebView2 on Windows, separates untrusted pages from its privileged interface, and publishes where those boundaries stop.
Permission belongs where the action happens
The most useful part of Zephium's design is not a model picker. It is the moment an agent crosses from reading to acting. Work asks before it reads history, notes, open tabs, or a folder. Site access can be set to Always, Ask, or Never for a particular site. Sending, posting, merging, deleting, or saving is documented as a prepare-then-approve operation. Banks, payment services, and health sites always require a prompt.
Those controls still demand judgment. Zephium explicitly warns that approving a command decides when it may run, not what the operating-system command is capable of doing. Its privacy documentation also says the browser profile is protected by the computer account and disk encryption rather than by separate profile encryption. Windows builds are not yet code-signed, and the project describes itself as beta despite its stable-numbered release. These are useful disclosures, not details to wave away.
A careful trial should use a disposable profile and low-value accounts:
- Ask the agent to compare several public pages and inspect every cited source.
- Give one site
Askaccess, then confirm exactly when the browser pauses. - Let it draft a harmless form submission, but stop before approval.
- Repeat with AI disabled to verify which ordinary browser features remain.
This is actual usage evidence in the practical sense: a reproducible evaluation of documented behavior, without treating catalog discovery or installation events as proof of adoption.
Three ways to share a browser with an agent
BrowserOS neo draws the boundary around an external agent. Its official material describes a local Chromium browser driven over MCP by clients such as Codex, Claude Code, or Cursor. It uses the accounts already signed into the browser, displays each step live, and keeps sessions, screenshots, and replays on the machine. That makes it attractive when the agent already lives in a development tool and needs a real authenticated browser. It also raises the central risk plainly: the agent is acting with the authority of those signed-in accounts. A separate browser profile and narrowly privileged accounts are sensible prerequisites, not optional polish.
Opera represents the mature-browser end of the spectrum. Its AI sits beside ordinary browsing and can summarize, compare, translate, or draft from page context. That is valuable when assistance matters more than autonomous execution. Opera's broader feature set and established daily-browser model reduce the need to adopt a new workspace, but integrated page-aware AI should not be confused with a carefully scoped action protocol. Users still need to understand what context leaves the browser and which service receives it.
Zephium tries to combine both ideas: a daily browser and an agent workspace designed together. Its canvas keeps results such as comparisons, tables, plans, files, and drafts as objects rather than a long chat transcript. BrowserOS neo instead makes the browser a controllable tool for an outside agent. Opera keeps AI closer to an assistant embedded in a conventional browser. None is universally safer; each places review, identity, and model context at a different layer.
Choose the trust boundary before the feature list
The deciding question is not which browser can click a button. It is who holds the session, who sees the page, and where approval occurs. Before adopting any agent browser, map the signed-in accounts it can reach, the model provider receiving content, the local files it may read, and the actions that require confirmation. Test recovery after a canceled run, a network failure, and a mistaken instruction. Confirm that logs and replays do not become a second store of secrets.
Agent browsers can make web work legible instead of hiding it in automation infrastructure. Their value depends on preserving the human's ability to inspect, interrupt, and refuse. Zephium's contribution is to make those handoffs part of the browser's visible workspace. Its youth means the right posture is evaluation, not automatic trust—but the experiment is concrete enough to run and the boundaries are documented well enough to challenge.