Build like the model will eat your harness in a year

Sonya Huang, hosting Sequoia's *Training Data* just after Google I/O, brings on Logan Kilpatrick, who runs Google AI Studio and the Gemini API and spends his days on the next generation of builders. Sundar Pichai had opened I/O by calling this the agentic Gemini era. Kilpatrick's first correction is timing. Google had already talked this way around Gemini 2.0, which felt early. The Gemini 3.5 moment, he says, is when agentic coding and agentic products actually started to feel true.

Historically, Google's roughly 50 products had no single through line. Gemini became the first: everything used the model in some way. A second line is now forming. The Antigravity agent harness, announced at I/O, is the layer other Google products are rebasing onto so they can take actions for users instead of only answering them. Antigravity is not just an IDE. It is a core IDE, a web agent-first experience, a CLI, and an SDK — an ecosystem designed to meet developers where they already work, including a managed agent through the Gemini API if you do not want to own the infrastructure.

The same harness, he says, will power agent features in Search, in the Gemini app, across Cloud, and in AI Studio. Huang tries the summary: the language model used to be how AI got into every Google product; now the coding harness is that connective tissue. Kilpatrick tightens it. Coding is a specialized use of an agent harness, but coding has also proved to be the general-purpose agent harness. About 80 percent of the stack is shared. AI Studio specializes it for vibe coding; the Gemini app specializes it for a consumer, always-on, 24/7 agent.

Huang's sharp question is cannibalization. If an agent answers your email, do you still open Gmail? More agency could mean fewer human eyeball-hours on Google's surfaces. Kilpatrick reaches for search. At the start of the current AI era, everyone assumed question-answering would be negative-sum for search. It was not. People searched more, and a second market appeared at the same time: agents searching. Human time in the world is finite, he grants, and the picture is clearer for the next one to two years than for three to five, when the products will look different. For now the ecosystem still feels positive-sum.

The deeper shift is the scoreboard. He talks with Demis Hassabis about this constantly. The point of the technology is that it goes and does the thing for you. Success for Google, in that frame, is not maximizing time in front of Google products. It is maximizing outcomes so customers can go live their lives. Huang has a phrase stuck in her head: agent-led growth. She already lets coding agents pick the database. Shopping, she suspects, will follow, and then advertising and value capture have to change. Kilpatrick's guess is milder than the panic. SEO and GEO — generative engine optimization — are highly correlated proxies stacked on top of each other. The shift, he thinks, will compound rather than snap.

On a crawl-walk-run scale, Google's suite is crawling, and he says that out loud. The company has on the order of 13 products with a billion-plus users. Stewardship matters. The long tail of customers is not ready for AI to run everything; they want to stay in the driver's seat. Search is the quintessential careful case: you bring people along, you do not rewrite how they use the internet overnight. Closest to walking are the Gemini app and Spark, with a 24/7 agent that might actually take actions, and Antigravity, where autonomous coding agents can chew through billions of tokens and thousands of dollars — even rebuild operating systems. Google DeepMind is taking the frontier look; the rest of Google is getting there incrementally, which he finds reasonable.

Will Google end up with three AI surfaces or ten thousand? Kilpatrick thinks humans like compartmentalized products. A calendar app that just shows the calendar beats a general-purpose blob that makes you spend mental energy steering it. Generative interfaces sound cool and often add cognitive overhead; Huang's hot take is that slide decks survived because the information sits in a stable place. He does not know Google's eventual count. For the ecosystem, he expects more products, not fewer. Whether Google itself consolidates to three or lives with thousands is a strategic choice about what customers will tolerate.

When Huang talks to enterprises, they say agents only really work in coding. Kilpatrick's answer depends on the bar. If the domain has not crossed a quality threshold, offloading hard work will fail. What he wishes someone would publish is not only tokens consumed — OpenRouter already sketches how much more intelligence is in the world — but how long the average agent run lasts. He does not think they ship that number, but he suspects they have it. Each new model drop spikes the curve. Labs talk about three days of autonomous work; that is the extreme. The practical curve is still low and rising fast. Enterprises that have not felt agents outside coding, he says, will feel it this year.

Long-horizon agents are a DeepMind KPI, and coding agents especially, because a great coding model is an accelerant for every other bet. Then Huang asks the uncomfortable question. Her developer friends lived on Claude; OpenAI declared a code red; Codex caught up; the split is now roughly 50/50 Claude and Codex. She does not hear much Gemini. Kilpatrick adds a chapter she skipped. In December, after Gemini 3 landed, the narrative was that Google had won. Then, over the holidays and into January, attention snapped to agentic coding. The speed of the narrative flip, not only the flip, is what he wants remembered. Things have been in warp speed since.

He does not call her observation unreasonable. Behind the scenes, Google is pushing the coding frontier as fast as it can. The hard lesson is that it is very difficult to make a great model for long-running software work if you do not have a product that does that work. That is why the Windsurf deal happened, why those people came over, and why they built Antigravity. Sundar showed the internal token-consumption curve at I/O; the engine is spinning, and model progress still takes time. Kilpatrick calls the coding group the Avengers of AI inside Google. Gemini 3.5 Flash is already a down payment: a Flash-tier model that, on coding, beats every Pro model Google had released, with those gains coming entirely from post-training.

Outsiders also forget pre-training windows. A lab can look behind while a large cluster run is still in flight, and pre-training has historically been a DeepMind strength. On dogfooding he is not precious. He uses every model and every product because otherwise you cannot grok the ecosystem; Gemini still has to be the daily driver. Google has a hundred thousand-plus engineers who can feed the flywheel with A/B tests and live experiments. As for a soft takeoff — a good enough agentic coding model accelerating research — it seems obvious to him, which he worries is Kool-Aid. On the model side, humans still allocate 10,000 TPUs; nobody will accidentally start a nonsense job. On the product side the signs are already there. His team is shipping mobile apps with Antigravity faster than any Google team he knows of, and Josh's team delivered the Gemini macOS app end-to-end faster than any Mac app in the company's history.

Kilpatrick has said before that a system which can build anything in code is narrow superintelligence, because humans cannot compete at that level. Asked whether we are there, he says it kind of feels that way for coding right now. Quantifying it is messy. The important fact, which he wants to let wash over him, is that being this good at code is already enormously impactful even if the model is not yet great at everything else. Obsessing over AGI in a way that erases present capability is a bad trade. He tries to hold both: build general-purpose technology, and do not talk past how useful this already is.

The surprise is agency. Coding agents have not taken it away from human developers; they have increased his. Ideas that used to sit slightly out of reach — *wouldn't it be nice* — now invert. He kicks an idea around and wonders whether he should make it more ambitious. The new burden is that an MVP feels insufficient. The technology lets him go ten steps further, so he has had to reset his own level of ambition. He expects the same pattern in other verticals, arriving as jagged superintelligence: sharp in some places, flat in others, a pile of narrow peaks before anyone can claim the general problem is solved.

The next peaks, he thinks, will be domains with better verifiability: math, finance, maybe science. He wants those positive, visible wins early in the sequence so the public story of AI is not only a race up SWE-bench. He is not a mathematician, but the current burst of work on proofs fits. Huang reads him an old tweet: last October he said everyone would be able to vibe-code video games by the end of 2025. Close, he says, and obviously not AAA — not the next *Call of Duty* or GTA. Three.js unlocked a lot; sprite generation and other game-native jobs still have rough edges a coding agent does not solve. You need orchestration, reliability, reusable depth, and taste.

Early AI Studio data had games at about 20 percent of apps. That is no longer the top category. Roughly 20 percent is now finance-related, much of it crypto, then personal productivity and generative media. DeepMind still cares: Demis Hassabis came into AI through games, and the Kaggle team built a game arena with GDM as a proxy for AGI progress. A random person with a good idea vibe-coding a truly fun game? He wants to say this year. Model quality is not the gap. The gap is someone who knows games putting the scaffolding together, plus discoverability — people do not even know they can — and a few capabilities that may be weeks or months off. Short term, coding agents plus a game engine beat a pure world model. World models are too open-ended until someone grounds them; that grounding is itself a startup.

World-model definitions, Kilpatrick says, are blurring, and Omni — launched at I/O — is the exhibit. Any input, any output. Hassabis framed it as a world model because of how much of the world it understands. Architecturally it is not the old Genie-style, action-conditioned video model, which was famously expensive and unscalable. Huang restates the shift: "world model" now often means a model with world understanding, not strictly an action-conditioned video generator. Omni is not real-time, but it can already produce many of the same visual results, which is why he thinks the video-model and world-model stories will collide in a less obvious way.

Under the hood, the point he wants remembered is that it is a single model. Google used to train a suite: Gemini for text, Lyria for music, Nano Banana for images, Veo for video, plus a stack of audio models. Routing among eight models would have been the cheap fake of "Gemini Omni." This is a true Omni. It is the first crank of an Omni Flash turn, starting with the use case that currently works best — video editing — because the other modalities are technically functional but not yet state of the art. More capable versions are coming.

He keeps returning to a live demo. On stage with Tulsi, who leads the model team, someone in the crowd edited a talk video with Omni in real time. A dog walked on, other guests looked down and chuckled, it jumped into Kilpatrick's lap, he acknowledged it, kept talking, and petted it. The model caught the social timing. Huang is bullish on generative media for a podcast where visuals compete with content. Kilpatrick, for his own posts, still refuses AI stand-ins: his words, his voice, his face. Authenticity is alpha. What he likes about Omni is that it does not replace the person. It changes the set and the coffee table — the bits that were never him — which feels healthier than a world of AI avatars and island videos.

On the coding side, AI Studio now lets people write Android apps, a strategic bet that builders should meet other Google surfaces without nine different UIs. Kilpatrick built his first Android app there, a gardening tool for planting trees in the backyard, not a crypto client. He has not seen a vibe-coded breakout in the App Store that he can name, but the volume stunned him. Checking numbers that morning, they were at about 350,000 Android apps built in AI Studio since the previous week — software that mostly would not have been built at all, a lot of it personal. Native phone capabilities, and the richness of OS texting versus any AI chat app, are why Android feels to him like it is becoming a platform for builders, even if the web is already powerful.

Then Huang asks him to unpack the line in the title: the model eats the harness. Two years ago, he says, a model was a set of weights — tokens in, tokens out. We still say Gemini 3.5, GPT, Claude, but the thing under the name is an expanding system around the weights: agentic tool calling, hosted search and code execution, models spun up in containers, an agent harness. Scaffolding stays a few steps ahead of what is baked into the model; then the model eats it and the capability becomes native. External search and code-execution tools still matter because different providers fit different jobs. The harness is the quintessential current example. Everyone is saying the alpha is in building your own. Kilpatrick's 12-month view is that models will have digested that layer, upstreamed it, and moved the alpha somewhere else. Spinning your own harness then will look like installing a flashlight app on a flagship phone.

Huang's objection is lock-in: application companies build their own harnesses so they are not stuck with one lab. That logic, he says, holds while models are weak. A truly general model should be able to use someone else's harness. He has been arguing for a Harness Bench that measures how well each model adapts to unfamiliar harnesses. Over time you would expect all of them to manage every harness; if you are completely out of distribution, you are out of distribution on your own harness too.

So how do independent companies live? Both things feel true to him. There has never been more opportunity, and the models do more than they ever have. Capability overhang is itself a pile of alpha. Model labs chase general problems; vertical teams with domain knowledge, customers, and an ecosystem can run laps around them because focus is the startup superpower. Google has many products and an obligation not to focus on one domain. Twenty-four months ago people feared a smaller opportunity set for startups. The opposite happened. Coding closed the gap with companies that already had huge codebases, and the agentic primitive is a new product category. Firms willing to take more risk can win the users who also want to take risk.

Inside Google DeepMind the mood, he says, is a moment worth absorbing. The portfolio is strong, and you still watch another lab pull ahead in a pocket where Google under-invested, then watch the close-the-gap reflex. He has watched Hassabis's documentary *The Thinking Game* several times: get smart people in a room and solve the problem — that strike culture is still there. Leaders leak into culture. Hassabis is a Nobel-winning scientist; you feel that. Sam Altman is among the world's best businessmen; you feel that at OpenAI. Dario Amodei, from the outside, seems more esoteric, and Anthropic feels that way too. Hassabis started the mission to solve disease. It is easy to get lost racing SWE-bench. Kilpatrick's favorite *Silicon Valley* line is Gavin Belson's: we cannot let other people make the world a better place more than we can. Framed that way, the race is goofy and not zero-sum. DeepMind's Twitter bio now calls it the engine room of Google. Deploying Gemini into billion-user products is a problem only two companies have, and Google has 13 of those products. He does not get every tweet approved. Magic watered through too much process, he says, stops being the story.