Google Lets AI Be a Middle Manager: Chief of Stuff Can Delegate Work on Its Own, but the Name Is Still Inconsistent
TestingCatalog has revealed Google Antigravity’s upcoming agent: Chief of Stuff will support Gemini Live and be able to operate autonomously, spawn subagents, fix CLs, and generate rules from history. It is not yet public, and commenters are already worried about file conflicts and it learning bad habits.
A TestingCatalog tweet says that an AI agent called Chief of Stuff is about to appear in Google Antigravity and will get Gemini Live support. Its current setup is that it can operate Antigravity by itself and spawn subagents to carry out subtasks. The tweet notes: not yet open to the public. Source: @testingcatalog, with thanks to @lberondroth.
The tweet lists four specific capabilities: check conversations from the past 3 days that need user attention; check CLs that need attention and fix them directly in the corresponding conversation; learn from conversation history to generate a set of personalized rules for future agent interactions; and recommend potentially useful skills and plugins based on history.
CL can be understood as a changelist, common in code review workflows. For developers, the second item looks most like time-saving work: half the time is spent on small reviewer comments, and one commenter said they want to hand such fixes directly to a subagent. NoRulers’ comment was more direct: spawning subagents is now standard, but personalized rules are the part he would actually use.
The questions in the comments are also practical. Aarav Y asked whether it can spawn subagents across models or only with Gemini, and whether Antigravity can run unattended overnight. Cicero said spawning subagents is easy; the hard part is keeping them from stepping on each other's files. He once broke down a disk cleaner build: 8 agents each managed their own files, and another agent specialized in attacking all deletion paths. He cares about whether a subagent can be given a read-only review role. These permission and isolation issues are closer to real deployment than “being able to assign work.”
There is also some confusion about the name. The TestingCatalog tweet text says Chief of Stuff, while the on-site article and video sidebar show Chief of Staff. User lei-knight specifically asked which one is the internal codename. So far both spellings have appeared in public information, with no unified version. Stuff sounds more like a joke, while Staff sounds more like a serious job title.
Some people are also uneasy about “learning rules from history.” andrew said, “until it learns your bad habits.” Sachin said half of his conversation history is telling the agent it is wrong, and he doesn’t know what rules could be learned from that. This concern is reasonable: history contains corrections, trial and error, and temporary instructions, so directly summarizing it into long-term rules needs boundaries.
Kx’s comment was a bit self-deprecating: agents delegate work to agents, and in the end the only one with no work is the human. ShadowAguy described it as middle management finally being automated. Google has not yet made this agent public; whether the features will launch as-is, how Gemini Live will be integrated, and how subagent permissions will be managed all have to wait for the actual version. What is certain now is that Antigravity’s agent layer is moving in the direction of “finding work itself and dividing work itself.”
发布时间: 2026-11-02 09:51