An agent is a model running a multi-step task with tools. Google says Teamwork divides work among agents, including separate critics, and checks candidates with tools such as proof checkers and software tests.
Google reports that its LEAP system formalized an existing proof of the Weil conjectures in Lean, a language whose proof checker checks each logical step. It says the formalization used about 1.8 million lines of Lean across 28,800 subagent invocations.
The company describes human researchers framing problems and checking results. It says some Teamwork patterns are available in Antigravity and that more features will reach external users in coming weeks.
Caveat: Google’s account is a report of its own collaborations; a checked proof also depends on the formal statement matching the intended mathematical claim.
The author’s Show HN project uses Safari’s WebKit engine and exposes shell commands for opening pages, inspecting elements and clicking. Its dashboard groups tabs by agent workspace.
Named profiles share logins across workspaces. A secret-command option can fill a password without returning it in the assistant’s command reply, according to the README.
The project is MIT-licensed and requires macOS 26 or later. It documents a handoff for login, multi-factor authentication or CAPTCHA and a source build requiring Xcode 26 or later.
Caveat: It uses WebKit only; Chrome-specific sites or tools that depend on Chrome’s debugging protocol may not work.
(HN 74 points · 21 comments · project README)
Lovable application exposure study → STALE: the report says its figures update November 2 reporting, but its update date is unclear and the same URL was submitted to HN on November 2. A new HN discussion alone does not establish a new finding this cycle.
Language and AI energy study → STALE: Michigan’s account is dated November 5; the current discussion does not establish a new study result.