Agent Zero v1.11--v1.13: Visible Computer, ODF Documents, and Persistent Desktop

What Changed

Agent Zero shipped v1.11 (May 2), v1.12 (May 3), and v1.13 (May 5) as rapid follow-ons to v1.10's native browser introduction. These three releases finish the "visible computer" surface: multi-tab browser, LibreOffice desktop, ODF documents, and persistent desktop lifecycle.

Note: v1.11-v1.13 shipped in the days before this window (2026-05-07) and were not included in the prior finding (2026-05-07-agent-zero-full-computer-workcell, which focused on v1.10). The current window has no commits; these releases are covered here to complete the profile.

LibreOffice replaces Collabora (v1.11): The Collabora/WOPI document runtime was removed and replaced with a Markdown-first document store backed by LibreOffice, an Xpra virtual desktop gateway, and a persistent XFCE desktop session. DOCX, XLSX, and PPTX files now open in full LibreOffice sessions via Xpra. The Linux Desktop skill teaches Agent Zero to operate the XFCE/Xpra environment, including app launch, focus, click, cell editing, and stable folder entry points.

ODF-first document defaults (v1.13): ODT, ODS, and ODP are now the primary formats for Writer, Spreadsheet, and Presentation artifacts. OOXML (DOCX, XLSX, PPTX) remains available as an explicit compatibility option. This reduces vendor lock-in for document artifacts generated by agents.

Multi-tab browser with parallel fan-out (v1.11): The browser auto-registers tabs opened by sites (window.open, target=_blank, ctrl-click) and supports working across multiple tabs concurrently. A new multi action lets the agent fan out reads or coordinated mutations across tabs in a single tool call, with parallel execution for different browser instances.

Persistent desktop lifecycle (v1.13): A single Xpra Desktop iframe stays alive across canvas, modal, and keepalive hosts. Explicit shutdown is distinguished from crashes via a new shutdown/restart state and an XFCE panel "Shutdown Desktop" launcher (requires confirmation click). Unsafe logout, lock, and switch-user affordances are hidden. This changes the semantics of the desktop: it is now a persistent surface, not a session that resets on navigation.

Desktop state controls (v1.13): A new desktop_state helper and expanded desktopctl observe-act-verify commands give agents structured primitives for inspecting and controlling desktop state. Agent guidance now prefers structured/app-native/keyboard workflows over coordinate clicks: agents are instructed to use structured actions first and treat coordinate clicks as a last resort.

PTY file descriptor leak fixes (v1.12): POSIX PTY master descriptors are now properly closed when terminal sessions end, preventing /dev/ptmx exhaustion under sustained use. Closed or exited PTY sessions are detected before writes, with automatic retry/recovery.

Browser context recovery (v1.12): Cached browser contexts that have closed are detected before reuse. Stale pages, screencasts, and interaction state are cleared, and the Playwright instance restarts cleanly.

Operator Consequence

The LibreOffice desktop runtime makes Agent Zero's document surface genuinely powerful: agents can open, edit, and generate full LibreOffice documents in a real desktop session, not just read/write Markdown-wrapped representations. The tradeoff is system weight: the Xpra + LibreOffice stack is heavier than the prior Collabora integration.

The ODF-first default is a meaningful governance posture choice: document artifacts are now stored in an open format by default. Operators who need OOXML compatibility must explicitly request it. This reduces the risk of agent-generated documents being trapped in a proprietary format.

The persistent desktop lifecycle (v1.13) changes how operators should think about desktop state: it persists across interactions, which makes it more useful for long-running work but also means accumulated desktop state (open applications, changed settings, browser sessions) must be managed deliberately.

The "prefer structured over coordinate clicks" guidance is notable: it is agent guidance that steers behavior toward more inspectable, reproducible actions. Coordinate-based clicking is fragile (position changes between sessions) and hard to audit; structured actions like cell_edit and app_launch are more reliable and leave a clearer record of what was done.

Bitter Implication

Agent Zero is the clearest example in the watchlist of a "visible computer" workcell: the agent acts in a real desktop environment, browser, and file system, and the human can see what it is doing (screenshots, canvas, Xpra stream). This is the pattern BitterGrid should study for workcell design: bounded, inspectable, with visible state.

The "prefer structured over coordinate clicks" guidance models a useful principle for receipt design: actions that can be named and described are easier to record and audit than positional actions. A receipt for cell_edit(B3, value=42) is more useful than a receipt for click(x=423, y=187).

Signal

ODF-first document defaults and persistent desktop lifecycle are action-bearing:

• Operators running Agent Zero should verify that their workflows handle ODT/ODS/ODP output, or explicitly configure OOXML compatibility where needed.
• Operators running long-horizon desktop sessions should plan for persistent desktop state management -- the desktop no longer resets on navigation; cleanup is now explicit.