Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness
In a Nutshell
Logan Kilpatrick describes Google's shift from using the Gemini model as the connective tissue across products to deploying an "anti-gravity" agent harness that powers autonomous, long-horizon agents in coding, Search, and consumer apps. He argues the harness itself is the current source of competitive advantage because models will soon absorb most scaffolding, and that coding has emerged as the general-purpose proving ground for agentic systems. Key bets include measuring agent run duration as a core KPI, betting that narrow superintelligence in verifiable domains will compound into broader capability, and positioning DeepMind's scientific culture as the engine for Google's frontier products.
These notes were generated by AI and may contain inaccuracies.
Logan Kilpatrick shared an example of using Gemini to edit a video in real time during a talk with Tulsi, who leads the model team. Someone in the crowd took a picture and edited it with Gemini, adding a dog that appeared on stage. In the edited version, other guests looked down, saw the dog, and chuckled while Kilpatrick continued speaking and petted the dog. The model handled subtle nuances like the guests' reactions and the interaction correctly.
Logan runs Google AI Studio and the Gemini API, spending significant time thinking about and building for the next generation of builders.
Sundar opened Google I/O by calling this the "Agentic Gemini era." Kilpatrick explained that while Gemini 2.0 mentioned some agentic concepts earlier, the Gemini 3.5 era is when agentic products are actually becoming reality. For Google, this includes agentic coding and agentic products across the board.
The agentic layer is powered by the anti-gravity agent harness, announced at I/O. This serves as an additional through line connecting all Google products. Historically, prior to Gemini, there was no single through line across Google's 50+ products. Gemini became that through line, and now anti-gravity is emerging as the new through line as products rebase to become agentic native and take action on behalf of users.
Anti-gravity encompasses multiple components: a core IDE, an agent-first web experience, a CLI, and an SDK. It functions as an ecosystem designed to meet developers wherever they are. Developers can use it through the Gemini API for a managed agent without infrastructure work.
The same harness powering anti-gravity also powers agent functionality across other Google products including Search, the Gemini app, Cloud, and AI Studio.
Previously, the language model via the Gemini API served as the through line for embedding AI into Google products. Now the coding harness itself is being used across products, meaning the coding agent drives more agentic properties.
Kilpatrick clarified that coding represents a specialized use case of the agent harness, but coding has proven to be the general-purpose agent harness beyond just coding applications.
There is nuance between the two. The base harness contains approximately 80% of the same components, with specialization applied for different use cases. The harness used in AI Studio is specialized for vibe coding, while the harness in the Gemini app is specialized for consumer always-on 24/7 agents.
When asked about potential cannibalization from agentic capabilities, Kilpatrick noted that early concerns about AI answering questions negatively impacting search proved unfounded. Search has been positive sum, with people searching more. Agents are also searching, creating additional market activity alongside increased human search behavior.
Kilpatrick emphasized that the point of building technology is for it to do things for users so they can live their lives. Success for Google means maximizing outcomes for customers rather than maximizing eyeball time in front of products.
Kilpatrick discussed how agent growth changes advertising and value capture for aggregators. The trend shows proxies between different optimization approaches, such as SEO correlating with GEO (generative engine optimization). The shift appears less radical than assumed because these systems compound on top of each other.
Google's product suite is currently at the crawl stage of agenticness. The company has 13 billion plus user products, creating inherent product tension. Some labs experiences are closer to walking or running, but most product experiences remain at crawling due to stewardship responsibility for products used by many people.
Users want to remain in the driver's seat and are cautiously taking first steps. Search exemplifies this, with the team having responsibility to bring people along without fundamentally changing how they interact with the internet.
The Gemini app is closest to walking, with Spark featuring a 24/7 always-on agent that can take actions on behalf of users. Anti-gravity represents another frontier use case, with autonomous coding agents capable of rebuilding operating systems, processing billions of tokens, and spending thousands of dollars autonomously.
Google DeepMind takes a frontier approach, while the rest of Google's products advance more incrementally.
Kilpatrick believes compartmentalization and specialization of products serves human consumption patterns well. A single product doing everything inherently requires more mental energy and time to use. People prefer clicking a calendar app that simply shows their calendar without additional complexity. The idea of a generative interface sounds appealing but may create cognitive overhead.
Kilpatrick guesses the ecosystem will see many more products targeting different use cases. For Google, the strategic decision between having 10,000 products versus three comes down to what customers want.
When asked about the claim that coding agents are the only agents enterprises have seen working, Kilpatrick noted it depends on the bar for "working." For complicated tasks in domains where models haven't crossed quality thresholds, the statement holds. However, he pointed to OpenRouter measuring total token consumption as evidence of increasing intelligence deployment.
Kilpatrick is interested in measuring average agent run duration and task length. Model labs are announcing releases that accomplish three days of autonomous work, with signs of long-running tasks trickling up quickly.
Long-running agents matter as a KPI for DeepMind. The organization maintains a portfolio of different bets. Coding agents specifically matter significantly as an accelerant for every other part of the business.
Kilpatrick addressed why developers are split between Claude and Codex with less Gemini usage. In December, the narrative was that Google had won with Gemini 3 representing a profound improvement. The narrative then shifted to agentic coding developments over the holidays and into January.
Google is pushing the frontier on coding aggressively. Anti-gravity is important for this effort. The company recognized it is hard to make a great coding model for long-running suite work without a product that does that work, which is why the Windsurf deal occurred and those individuals built anti-gravity.
Sundar showed at I/O a graph of token consumption growth inside Google. The engine is spinning, though model progress takes time. Kilpatrick described the group working on code as "the Avengers of AI" internally, with some of Google's best people pushing this effort.
Gemini 3.5 Flash is better than any pro model previously released from a coding standpoint, despite being a flash model. This represents post-training gains, as 3.5 Flash achieved all improvements through post-training rather than pre-training.
Kilpatrick noted that external perceptions of being behind may miss context about big pre-training runs and cluster availability. DeepMind has historically excelled at pre-training with some of the world's best people.
DeepMind folks use other models because it is healthy for understanding ecosystem developments. Kilpatrick personally uses all models and products. However, everyone must use Gemini models for feedback flywheel purposes. DeepMind and Google have over 100,000 engineers using the models and giving feedback, which should be a competitive advantage through scale of engineering resources and ability to run A/B tests and live experiments.
Gemini serves as the daily driver for most people.
Kilpatrick believes the narrative about soft takeoff through agentic coding accelerating research progress is obvious but questions whether he has drunk too much Kool-Aid. Signs are visible from a product perspective. His team built mobile apps using anti-gravity and launched them faster than any team at Google has built mobile apps previously. Josh's team delivered the Gemini Mac OS app end-to-end faster than any previous Mac app delivery at Google.
Kilpatrick acknowledged that coding currently feels like narrow superintelligence. The capability works incredibly well for code, which is impactful enough on its own. He tries to hold two perspectives simultaneously: building AGI remains important, but current present capabilities should not be overshadowed by AGI development focus.
The positive outcome is that this capability has not taken away from human developers. Kilpatrick personally feels more agency as a human developer and can tackle more as a result.
The speaker notes feeling capable of tackling more ambitious problems than before. Previously, ideas would feel slightly out of reach and remain aspirational, but now the opposite problem exists where ideas can be made even more ambitious because the technology enables it. This shift creates a different layer of burden because the minimum viable product approach no longer suffices, requiring going ten steps further instead.
Resetting ambition levels is something the speaker has spent significant time thinking about. This pattern is expected to occur across vertical super intelligence domains, resulting in what the speaker describes as "jagged super intelligence" before achieving broader solutions.
Domains with better verifiability are expected to see gains happen more quickly. Math, finance, and science are highlighted as particularly interesting areas where verifiability could drive rapid advancement. The speaker emphasizes the importance of positive, impactful applications happening as early as possible in AI development so people understand the technology's potential benefits.
Recent developments in math proofs are noted, though the speaker acknowledges not being a mathematician. A tweet about Ardos having problems is referenced as a memorable observation.
A tweet from last October predicted that "Everyone is going to be able to vibe code video games by the end of 2025." The speaker assesses this as feeling close to reality, though AAA games like the next Call of Duty or GTA are not yet achievable. The interesting aspect of video games involves needing to build supporting infrastructure beyond just models.
3JS is cited as an example of tooling that enables previously impossible things, but rough edges remain that coding agents alone cannot solve. Sprite generation is identified as an area where models are not yet strong natively, requiring orchestration layers and tooling. Other core gaming elements need high reliability, requiring product scaffolding work to create reusable, replayable experiences with depth and taste.
Early data from AI Studio showed approximately 20% of all apps being games. While games are no longer the most popular category, significant game development activity continues. Finance-related apps, particularly crypto projects, represent around 20% of usage. Personal productivity tools and generative media applications are also popular categories, with Google's generative media suite being noted as particularly strong.
DeepMind maintains deep roots in gaming, with Demis Hassabis having started his AI career in this domain. The Kaggle team collaborates with DeepMind to build a game arena as a way of testing progress toward AGI using games as a proxy. This approach is deeply connected to DeepMind's historical focus on games.
The speaker predicts that a random person with a good idea could vibe code a really fun, playable game this year. Model capability is viewed as sufficient, with the gap lying in product implementation rather than model quality. The challenge involves someone who understands what makes a great game assembling the right scaffolding. Some people are already doing this successfully, though discoverability and awareness remain barriers. Certain model capabilities are weeks or months away from crossing the necessary threshold.
The definition of world models is expected to blur over time. Coding agents will increasingly resemble world model type systems, but scaffolding remains necessary to make world models useful for real applications. Startups are working on figuring out scaffolding that takes world models from open-ended spaces and grounds them in tangible, recurring use cases.
The inherent nature of current world models makes them less well-suited for games in their present form. However, progress has been rapid, and this assessment could change within two years. In the short term, coding agents combined with game engines are expected to deliver more value for game development.
Omni is presented as an example of evolving world model definitions. Launched at IO, it accepts any input and creates any output. Demis framed it as a world model due to its level of world understanding. Technically, this differs architecturally from previous world models, which is viewed positively as it may prove more scalable. Traditional online world models like Genie have been very expensive to run.
Traditional world models are conceptualized as action-conditioned video models. The current meaning of "world model" refers to a model with some understanding of the world rather than being strictly an action-conditioned video model. The line is blurry because these models can perform many of the same use cases that would be described or created visually with traditional world models, though not in real time currently.
Omni is a single model rather than routing to multiple specialized models. Historically, separate models handled different modalities: the baseline Gemini text model, audio models, music models like LIA, video models, and a suite of audio models. The goal was to create a single model capable of handling all these functions rather than maintaining separate systems.
The first available use case is video editing capability. Other functions are technically operational but quality is not yet state-of-the-art. This represents the first iteration called Omni Flash, with more capable versions planned.
A demonstration involved editing video in real time during a talk with Tulsi, who leads the model team. A dog appeared on stage in the original footage, and the edited version showed other guests looking down, chuckling at the dog, with the dog jumping onto the speaker's lap. The speaker acknowledges the dog while continuing to speak and petting it. The model handled the subtle nuances correctly, demonstrating world understanding in content creation contexts.
The speaker has historically avoided using AI to create personal content, preferring authenticity with original words, voice, and image. Omni is appreciated because it changes elements around the person rather than the person themselves. The set and coffee table can be modified while keeping the original words and personhood intact. This aligns with a preferred vision of generative media that avoids AI avatars.
Android app development capability was launched in AI Studio. The strategic principle involves bringing Google ecosystem capabilities into AI Studio so users can access different parts of Google without navigating multiple UIs. Android enables people who otherwise would not have built Android apps to do so. The speaker built a gardening app as their first Android app in AI Studio.
Approximately 350,000 Android apps have been built in AI Studio since the previous week. Many of these are personal applications that likely would not have been built otherwise. The ability to unlock native phone capabilities is viewed as particularly interesting given the distributed context available on devices. Android is becoming positioned as a platform for builders.
Web platforms are powerful, but operating systems provide native richness that cannot be unlocked through web alone. Text messaging experiences across major operating systems feel richer than AI chat apps, and the speaker would prefer AI integration within existing texting apps rather than requiring separate applications.
The principle that the model eats the harness or scaffolding is viewed as true. Two years ago, models were understood as sets of weights focused on sending tokens in and getting tokens out. Now models encompass entire expanding systems built around weights, including agentic tool calling, hosted tools, search, and code execution. Models run in containers with agent harnesses.
Scaffolding is often steps ahead of what is baked directly into models. Eventually, models digest that scaffolding and incorporate it natively. External scaffolding retains value in certain cases, with search and code execution cited as examples where flexibility in providers remains desirable.
The agent harness is identified as the quintessential current example. Building harnesses is currently where alpha exists, but this is expected to change within 12 months as models upstream this functionality natively.
Application companies build their own harnesses partly to avoid lock-in from model provider harnesses. However, as model capability improves, this becomes less true over time. A generalized model should be capable of using other harnesses. The speaker proposes creating "harness bench" as a benchmark measuring how well models adapt to different harnesses.
Despite models doing more than ever before, significant opportunity remains for independent companies. The capability overhang represents substantial alpha. Model companies pursue general problems, leaving value in verticalized domains where domain expertise, customer knowledge, and ecosystem understanding provide competitive advantages. Focus is described as the superpower of startups.
Large companies like Google have obligations to address many products and user bases, preventing deep focus in single domains. Startups can focus intensely. The opportunity space has not contracted as some predicted 24 months ago. Coding capabilities help close gaps with larger companies by enabling faster software development. The agentic primitive represents a new category for product development, with varying risk appetites across companies creating different market opportunities.
Logan Kilpatrick describes the current moment at Google DeepMind as incredible, actively trying to take it all in and reflect amid the chaos of so much cool stuff happening. He shares three observations about GDM's culture. The first centers on focus from a portfolio perspective. While GDM has one of the strongest portfolios, other labs sometimes pull ahead in specific areas where DeepMind has underinvested or not focused enough. He appreciates watching how the team works to close those gaps.
Kilpatrick references watching the Demis thinking game documentary multiple times and notes how the original culture persists today: getting a bunch of smart people together to solve problems. This approach remains very similar to the early days. The second observation is how culture permeates from leadership. Demis, as a Nobel Prize scientist and one of the original figures in the field, shapes DeepMind's deeply scientific culture. By contrast, Sam represents one of the world's best businessmen, which shows in OpenAI's culture. Dario at Anthropic appears more esoteric, and that quality seems embedded in their company's DNA. The other labs are interesting, but Kilpatrick values DeepMind's very scientific approach to the world.
Demis started the mission literally to solve disease and other real human problems. It's easy to get lost in the competitive race of pushing numbers higher on benchmarks like SWE, but the actual purpose is solving problems humans face. Kilpatrick's favorite quote from Silicon Valley is the Gavin Belson line: "we can't let other people make the world a better place more than we can." He thinks about this often, noting that when framed this way, the competition seems really goofy. The work is very much not zero-sum.
The third cultural observation is that DeepMind functions as the engine room of Google, which is now literally the Twitter bio of the DeepMind account. On one hand, there's the deep-rooted lab culture; on the other, there are partners across the entire Google ecosystem including Android, Google Cloud, Gmail, and Workspace. It's an interesting blend of research work and applied work with forefront customers. Deploying Gemini to billion-user products is a problem only two companies in the world face, and Google has 13 of those products. This kind of innovation can only happen inside Google.
When asked whether joining and tweeting frequently caused heartburn at Google, Kilpatrick explains he gets sign-off from comms. One silver lining of his Google experience has been how great the marketing and comms teams are to work with. Their job is to protect Google and tell the right story, and he has tremendous appreciation for that partnership. He values being able to tell stories that resonate with developers in an authentic way without needing every tweet approved. He tries to walk the line of not burning the trust and goodwill he's accumulated.
It's really hard for a big company like Google to tell an authentic story because of the number of people and opinions involved. The magic of Google gets watered down through process, and the beautiful story—that Google is doing the most interesting technology in the world and helping users with the hardest problems—gets missed. He feels it's a privilege to help tell that story.
The interviewer expresses appreciation for the sincere human touch Logan and Josh have brought to the most important problem of the time. Logan thanks the host for the far-ranging conversation covering agents, coding, world models, harnesses, and GDM culture. He expresses excitement about seeing what people build with the topics discussed and playfully hopes a dog appears to fulfill his dog dreams.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
