Google DeepMind’s Logan Kilpatrick: Why the Model Eats the Harness

The entire startup ecosystem is racing to build agent harnesses. Logan Kilpatrick, who leads Google AI Studio and the Gemini API, argues that scramble has a roughly 12-month shelf life. Models will absorb the scaffolding and run it natively, so the edge moves elsewhere. Google’s own bet runs in parallel: a single agent harness, born from the Windsurf team and now called Antigravity, has become the connective tissue across search, the Gemini app, Cloud, and AI Studio — the role Gemini-the-model used to play. Logan makes the case that coding already feels like narrow superintelligence, and that “jagged” vertical superintelligence (in math, finance, and science) will arrive well before AGI. He argues Google’s real goal is maximizing outcomes for users, not eyeball time. He unpacks Omni, the single model built to replace multiple separate systems Google once trained for text, audio, music, image, and video. His throughline: AI is an accelerant for human ambition, not a substitute for it.

— Sequoia Capital YouTube channel (hosted by Sonya Huang)

A ~51-minute interview on Sequoia Capital’s Training Data podcast (11 June 2026) with Logan Kilpatrick, who runs Google AI Studio and the Gemini API at Google DeepMind, hosted by Sonya Huang (Sequoia Capital). The conversation ranges across agentic AI, AI coding, world models, and DeepMind’s culture, but its load-bearing contribution to the wiki is the “model eats the harness” thesis — a named, from-inside-a-frontier-lab counterpoint to the wiki’s agent-harness corpus.

This is also the first source acquired through the new zotero-acquire channel (Zotero item R4UDMW3G, collection ai-wiki); the Zotero stub carried only metadata + the abstract, so the transcript was re-acquired via the youtube-transcript-skill per the YouTube-URL rule, and the Zotero provenance was merged into the raw file’s frontmatter.

TL;DR

• “The model eats the harness / scaffolding.” The “model” is no longer a set of weights — it is “an entire expanding, sprawling system built around the weights” (tool-calling, hosted tools, search, code execution, containers, the agent harness). Scaffolding runs “a couple of steps ahead” of what’s baked into the model; then the model digests that scaffolding, it gets upstreamed into the model, and the alpha moves elsewhere. Kilpatrick: “everyone’s like, the harness is where the alpha is, and I think that perhaps won’t be true — at least in the way we think of the harness today — in 12 months.”
• The lock-in counter-argument, answered. Asked why builders roll their own harness (to avoid model-provider lock-in / keep flexibility): “that starts out perhaps true, but as model capability improves it becomes less true over time… you don’t have a generalised model if it can’t use another harness.” Proposes a “harness bench” — an ecosystem benchmark for how well models adapt to different harnesses.
• Antigravity = Google’s single agent harness. Born from the Windsurf team/deal. It is “a lot of things” — a core IDE, a web agent-first surface, a CLI, an SDK, and a managed agent via the Gemini API. The same harness powers search, the Gemini app, Cloud, and AI Studio — the new “through line” across Google’s ~50 products, replacing the role Gemini-the-model used to play. “Coding has proved to be the general-purpose agent harness.” ~80% shared base, specialized per use case (vibe coding vs. a 24/7 consumer agent).
• Jagged / narrow superintelligence. Coding already “kind of feels like narrow super intelligence.” Kilpatrick expects “jagged superintelligence” — vertical domains with verifiability (math, finance, science) crossing the bar before AGI — and wants the high-positive-impact verticals to land “as early as humanly possible.”
• Outcomes over eyeballs. On cannibalization (agents reduce human eyeball-hours): AI answering questions was assumed negative-sum for search but turned out positive-sum (humans and agents search more). “Success for Google probably doesn’t look like maximizing eyeball time… it’s maximizing outcomes for customers.”
• Agenticness is at “crawl.” Most of Google’s 13B+-user products are at crawl (stewardship); the Gemini app is closest to “walk”; frontier autonomous coding agents (Antigravity) can spend thousands of dollars / billions of tokens. Long-horizon agent run-length is a rising KPI; “a great coding model is an accelerant of every other part of your business.”
• Candid on Gemini vs. Claude/Codex. Devs are ~50/50 Claude/Codex and “I don’t hear a ton of them using Gemini” despite the December Gemini 3 leap. Why Antigravity matters: “it’s really hard to make a great coding model… if you don’t actually have a product that does that” (the Windsurf rationale). 3.5 Flash’s coding gains were “all post-training,” beating prior Pro models; external “who’s behind” reads miss pre-training/big-run timing.
• Omni = one model replacing eight. A single any-input → any-output model (not routing) replacing separate systems for text (Gemini), audio, music (Lyria), image (nano banana), and video (Veo). Demis framed it as a world model (level of world understanding); the term “world model” is blurring. First iteration = Omni Flash, currently best at video editing. Near-term vibe-coded games = coding agent + game engine, not pure world-model-based (world models still need scaffolding to be grounded).
• Vibe-coding / the app layer. 350,000 Android apps built in AI Studio in one week — “apps that probably no one was going to build before,” many personal. “Building software to solve your personal problem is very real right now.” Most popular AI Studio categories: ~20% finance/crypto, games (was ~20%), personal productivity, gen media. Android “becoming the platform for builders.”
• Startup edge despite “model eats the harness.” “Both of these things feel true” — models do more than ever and “there’s never been more opportunity.” Capability overhang + verticalization + focus as the startup superpower (big companies “can’t focus”); coding closes the gap vs. incumbents’ established codebases; the agentic primitive is a new product category.
• DeepMind culture. “The engine room of Google.” Strong portfolio that occasionally underinvests and then closes gaps; culture permeates from leaders (Demis = Nobel scientist → scientific approach); explicitly non-zero-sum (the Gavin Bellson “make the world better” joke); deploying Gemini to billion-user products is something “only two companies in the world” can do — Google has 13 such products.

What was actually ingested

Full auto-generated transcript, ~51:09 / ~1000 ASR lines, with light cleanup; chapters preserved as section headings (16 chapters from Introduction → Inside DeepMind Culture). The opening ~80 seconds is off-camera banter about an Omni video-editing demo (a dog edited onto the stage in real time), revisited at 31:10. No sponsor breaks. The transcript is an ASR snapshot; proper nouns (Antigravity, Omni, Lyria, Veo, nano banana, Windsurf, Demis, Tulsi) were sanity-checked against the description and chapter titles.

How this source touches the wiki (dynamic capabilities)

• digital-sensing/digital-scenario-planning — the 12-month harness-absorption forecast, the jagged-superintelligence sequencing (verifiable verticals before AGI), and the soft-takeoff discussion are explicit future-state scenarios about where the AI frontier moves next.
• digital-sensing/digital-scouting — Kilpatrick continuously scans the competitive frontier (Claude/Codex/Gemini split, OpenRouter token-consumption as a market signal, capability-overhang, “harness bench” as a measurement gap the ecosystem should fill).
• digital-seizing/strategic-agility — Google rebasing all products onto the Antigravity harness, the Windsurf acqui-hire to get a coding product, and the outcomes-over-eyeballs pivot are concrete strategic-agility moves that reallocate the company around a new through line.
• digital-seizing/balancing-digital-portfolios — DeepMind’s “portfolio of bets,” the one-vs-many product-surface question (3 products or 10,000?), and the admission that a big company “can’t focus” are portfolio-balancing tensions.
• strategic-renewal/business-model — “maximizing outcomes, not eyeballs,” the agent-growth reframing of value capture, and the SEO→GEO advertising-model question are business-model-renewal claims at the level of how Google monetises attention.

Linked entities and concepts

• Concepts: agent-harness (the model-eats-the-harness debate), vibe-coding (AI Studio app explosion).
• Entities (promoted this ingest): Logan Kilpatrick, Google DeepMind, Sequoia Capital, Antigravity, Omni.
• Related org already in the wiki: Google (DeepMind is part of Google).
• Dangling (single-source mention, deferred): Sonya Huang (host), Demis Hassabis, Tulsi Doshi, Sundar Pichai, Windsurf (named as the team behind Antigravity — promote on second mention), Lyria / Veo / nano banana (Omni’s predecessor model lines).

Source-to-source relationships

• Contradicts 2026-05-04-rethinking-agents-harness-is-all-you-need — the sharpest edge. Both sources invoke the subtraction principle (harness components encode assumptions about what the model can’t do alone, and those assumptions expire as models improve). They draw opposite strategic conclusions: the Prompt Engineering source treats the (transferable, optimisable) harness as the durable reusable asset — “the reusable asset is not the model, it’s the harness” — whereas Kilpatrick treats harness alpha as transient, absorbed into the model within ~12 months. The disagreement is about durability of the moat, not mechanism.
• Supports 2026-03-05-lou-deepmind-autoharness-code-harness-synthesis — both are Google DeepMind framings in which the model subsumes the harness rather than humans hand-crafting a durable one. AutoHarness has a small model (Gemini-2.5-Flash) synthesize the harness as code, up to the harness-as-policy extreme where the LLM isn’t called at runtime; Kilpatrick generalises the same direction-of-travel to the model absorbing the harness natively. AutoHarness is the mechanism-level evidence; Kilpatrick is the strategy-level prediction.
• Other co-occurring agent-harness sources (Chatterjee, Kokane, Osmani, Karten/Zhang, Trivedy, etc.) share the topic but no typed edge added — Kilpatrick does not engage their specific claims, and the “model-eats-the-harness” thesis is already represented against them via the concept page’s Debates section. (Not every co-occurrence is a relationship.)