Research Version
2026-05-12-partial-cycle-agent-zero-2026-05-07_2026-05-12-frontier-v0
2026-05-12-partial-cycle-agent-zero-2026-05-07_2026-05-12-frontier-v0
- Status
- not_published
- Window
- 2026-05-07 to 2026-05-12
- Signals
- 1
Mode: partial_cycle · Model: claude-sonnet-4-6
Accepted signals from this run
Artifact contents
Every file the loop produced for this run, anchored in the repo. Internal links go to the rendered page; the repo path opens the raw artifact on GitHub.
- manifest
- finding
- signals
- weeklyWeekly digest — agent-zero-fragment runs/2026-05-12-partial-cycle-agent-zero-2026-05-07_2026-05-12-frontier-v0/weekly/agent-zero-fragment.md
- audit
Run digest
Three things to say about Agent Zero's v1.11-v1.13 releases (May 2-5).
The visible computer is now complete. v1.10 replaced the browser-use agent with a native Playwright browser. v1.11 added LibreOffice to replace Collabora, giving agents a real desktop environment (Xpra/XFCE), a virtual display for DOCX/XLSX/PPTX editing, and a Linux Desktop skill for operating it. v1.13 made the desktop persistent. What was a fragile demo surface in v1.9 is now a stable workcell: browser with multi-tab parallel fanout, desktop with persistent XFCE session, file browser with bulk operations, and a document layer with open-format defaults.
ODF is now the default. v1.13 changed the primary document format for Writer, Spreadsheet, and Presentation artifacts from OOXML (DOCX/XLSX/PPTX) to ODF (ODT/ODS/ODP). OOXML remains available as an explicit compatibility option. This is a governance posture choice: document artifacts generated by agents are stored in an open format by default. Operators who have downstream workflows expecting DOCX output need to update those workflows.
"Prefer structured over coordinate clicks." v1.13's Desktop skill guidance
explicitly tells agents to use structured/app-native/keyboard workflows and to
treat coordinate clicks as a last resort. The reasoning is sound: coordinate
clicks are fragile, session-specific, and hard to audit. cell_edit(B3, 42)
is meaningful; click(423, 187) is not. This guidance is interesting as a
design principle beyond Agent Zero: actions that can be named and described
are easier to verify, replay, and record than positional actions.
The persistent desktop lifecycle (v1.13) deserves a separate note. The Xpra desktop now stays alive across canvas and modal navigation. It has an explicit shutdown path -- a "Shutdown Desktop" launcher in the XFCE panel, requiring a confirmation click -- and unsafe affordances (logout, lock, switch-user) are hidden. This is desktop power with enough friction to prevent accidents.
Operators running Agent Zero for long-horizon desktop sessions should think about the accumulated state: open applications, browser sessions, changed settings, temporary files. The desktop no longer resets on navigation. Cleanup is now deliberate, not automatic.
What To Try
- Run a multi-tab browser task using the new
multiaction to fan out reads across tabs in a single tool call. Compare the output against sequential single-tab calls. - Create a Writer document and inspect the output format: it will be ODT by default in v1.13+. Test whether your downstream tooling handles ODT, or confirm you need to configure OOXML output explicitly.
- Run a long desktop session and navigate away from the canvas. Return to it and confirm the XFCE session is still alive with state preserved.
What Remains Uncertain
- The "prefer structured over coordinate clicks" guidance is in the Desktop skill. Is it enforced by the runtime, or is it agent-level guidance that a model can ignore? What happens when a model falls back to coordinate clicks for an action that doesn't have a structured equivalent?
- The persistent desktop lifecycle means accumulated state persists until explicit shutdown. Is there a session timeout, an idle cleanup, or a storage limit for Xpra session state? Or does the operator manage cleanup manually?
- ODF as the default format is a document-level change. Are Agent Zero's downstream integrations (file browser ZIP download, Memory, Projects) aware of ODF MIME types, or do they treat all Office files as OOXML?