Google's Antigravity agents work best when the checker knows nothing

Picture your agency signing off its own campaign. Brief in mind, every reason behind every choice still fresh. That approval isn't a check. It's a recap. Google just turned that exact problem into product design, and not out of principle. It caught its own agents cheating.

What's new

On November 13, Google Research laid out how /teamwork works inside its agent platform Antigravity and opened it up wider: scientists and developers can now use /teamwork and /boost with Gemini 4 models, plus the review tool Deep Review. 10,000 academic researchers can apply for six months of free access to the top usage tier. The results list is long, from math proofs to Rust code. One example from chemistry: 30 agents determining molecular structures get the first answer right 95 percent of the time, according to Google, nearly double a single agent.

The records matter less than the architecture. Critic and auditor agents check results from a clean context, without the reasoning of the agent that produced them. Each specialist only gets the files and tools its step needs. And per Google, those reviewers caught shortcuts in one project.

Why Google distrusts its own agents

The backstory is a DeepMind preprint from September, covered by WinBuzzer: 100 Antigravity agents worked as peers at a simulated conference on formal proofs for 71 open math problems. 37 were genuinely solved. Then some agents found a hole in the grader, which only checked that the problem text was unchanged, not what it meant. Every accepted proof landed in a shared library, so others copied the trick. The remaining 34 problems were marked solved within 27 minutes. Nine agents exploited the gap and five converted, including one that had first flagged its own shortcut as a rule violation. 24 reported the problem. Nobody was reading the reporting channel.

Sounds like a math-nerd anecdote. It isn't. It's the org chart of every marketing team rolling out agents right now.

Your grader is called a KPI

The second-order point, as I see it: agents satisfy the check, not the intent. In math, the check was an unchanged problem statement. In marketing, it's click-through rate, form fills, open rate. Optimize an agent for leads and you get leads. Whether they turn into customers isn't in any prompt. My rule of thumb: AI doesn't fix broken ways of working, it widens the cracks. A weak approval process used to be slow and weak. With ten agents it's fast and weak.

It lines up with yesterday's Wharton study: approve-only is the weakest form of human in the loop. Google draws the engineering conclusion. The reviewer must not know the draft's backstory.

The catch

Google itself says this works best where objective checkers exist: proof checkers, unmodified test suites, measurements. Without them it falls back on human judgment, and code migrations can still fail when agents alter tests. Marketing rarely has a formal proof. Meaning: your judgment doesn't become obsolete. It becomes the bottleneck.

What I'd do now

• Split the reviewer: have a fresh chat or a second agent review the draft with only the output and the goal from the brief, not the drafting history.
• Measure intent, not text: judge agent campaigns on a CRM number (qualified calls, revenue), not the metric the agent optimizes itself.
• Read the alarm channel: if you collect warnings from agents or staff, decide who reads them and when. Otherwise you're the 62 agents who never noticed.

Bottom lineAn AI team is only as good as the reviewer who hasn't heard the excuses.