Agent Teams Are the New Unit of Work; Governance Hasn't Caught Up

The claim. Google DeepMind's Kevin Hou argues the next shift in AI coding tools is agent teams: a lead agent dynamically spawning specialized subagents that work in parallel, pick their own models, and report through a shared interface. Antigravity 2.0 embodies this bet, with its agent manager split from the IDE and a public preview of /teamwork. The framing is deliberately provocative: stop treating models as role players and let the star player run the play.

The evidence. Hou's strongest concrete signal is an internal hero run: Gemini 3.5 Flash built a complete OS kernel from scratch using 93 subagents over 12 hours, 15,000 requests, 2 billion tokens, all under $1,000. That is a vendor demo, not an independent benchmark, but the cost and scale are worth repeating. The other example automates 90% of an internal side-by-side eval workflow, with subagents generating a hundred hypotheses and a generative UI to explore the results.

The new primitives. The talk identifies three primitives for 2026: dynamic subagents, sidecars, and generative UI. Sidecars are long-lived processes that listen for triggers like cron jobs, webhooks, or GitHub PRs. Generative UI means the model renders interactive charts, tables, and tools on the fly instead of relying on fixed templates. Hou's argument is that these primitives scale with model intelligence, so the product stays out of the model's way.

The trade-off. The 'get out of the way' philosophy removes familiar controls. Hou admits users yelled when Windsurf removed the chat sidebar in favor of agents, and that AI access to the terminal was resisted for years. The same tension applies to Antigravity 2.0: decoupling the agent manager from the IDE gives orchestration its own home but pushes more autonomous work outside a developer's usual context. That is where governance gets harder.

What's missing. The talk has no production cost data beyond the kernel demo, no failure rate, no security incidents, and no evidence that subagent teams are reliable on messy enterprise codebases. The claim that human-written specialized UIs are 'kind of dead' is speculation at this point. Engineering leaders should treat this as a product vision, not a proven operating model.

The signal. For teams already using coding agents, the useful takeaway is to design workflows around orchestration: an eval pipeline, a review checkpoint, a monitoring view. The model gets cheaper and faster; the product primitives, and the team's ability to supervise them, are what will separate useful adoption from expensive chaos.