Sequoia Capital · Published 2026-06-11

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

privateReview: unreviewedConfidence: high
agenticgeminiHarnessdevelopersmultimodalscaffoldingomnimodeldefensibilityevaluationcode-execution
Open on YouTube ↗

Summary

Overview

  • Speaker: Logan Kilpatrick
  • Channel: Sequoia Capital
  • Main topic: Agentic AI, Gemini API, and the integration of AI models into developer ecosystems and consumer products
  • Purpose: To explore the strategic and technical implications of agentic AI, model capabilities, developer tooling, and the evolving relationship between foundational AI models and application scaffolding. Logan Kilpatrick, head of Google AI Studio and Gemini API, discusses the transition of AI from simple chat interfaces to agentic workflows, the concept of the 'model eating the harness,' and the expanding scope of AI application layers across Google's ecosystem.

Topic Map

The Agentic Gemini Era and Antigravity

  • Explanation: Google's shift toward agentic AI, powered by Antigravity and Gemini 3.5, enabling models to take action on behalf of users across products.
  • Key claims:
    • Antigravity is the agent harness powering Google products.
    • The industry has entered the agentic coding and agentic product era.
    • Foundational language models are becoming the through-line for all Google products.
  • Examples:
    • Gemini app as an always-on 24/7 agent.
    • Antigravity powering search, cloud, and AI Studio agents.
  • Terminology:
    • Antigravity
    • agent harness
    • agentic coding
    • Gemini 3.5
  • Why it matters: It establishes a unified agentic layer across hundreds of Google products, transforming AI from a Q&A tool into an active participant.

Why the Model Eats the Harness

  • Explanation: As foundational model capabilities improve, specialized scaffolding and harnesses built around models are absorbed natively into the model architecture.
  • Key claims:
    • External scaffolding is frequently absorbed by advancing model capabilities over time.
    • Developers building custom harnesses risk being eaten by the model as intelligence improves.
    • True differentiation moves upstream as raw model capability increases.
  • Examples:
    • Custom coding harnesses vs. native coding agent capabilities.
    • Omni-model integration absorbing task-specific wrappers.
  • Terminology:
    • model eats the harness
    • scaffolding
    • agent harness
    • omni-model
  • Why it matters: It redefines startup strategy and product defensibility in the AI era, forcing builders to focus on domain-specific workflows and proprietary distribution rather than transient wrappers.

Google AI Studio and Developer Ecosystems

  • Explanation: The role of Google AI Studio and Gemini API in enabling developers to build agentic applications and mobile apps quickly.
  • Key claims:
    • AI Studio accelerates application development across platforms like Android and macOS.
    • Coding has proven to be the most robust general-purpose agent harness.
    • Scale of engineering resources at Google provides a unique advantage in deploying multimodal models.
  • Examples:
    • Building a garden planning Android app in AI Studio.
    • 350,000+ Android apps built using AI Studio.
  • Terminology:
    • AI Studio
    • Gemini API
    • coding agent
    • multimodal
  • Why it matters: It highlights how developer tooling lowers the barrier to entry for building production-grade AI applications.

Key Points

Maximizing Outcomes vs. Maximizing Eyeballs

  • Explanation: Google's product philosophy is shifting from keeping users on screens to maximizing user outcomes and task completion.
  • Evidence: Search and agentic tools driving successful task completion rather than raw ad impressions.
  • Practical implication: Product metrics must focus on user success and autonomy rather than mere engagement duration.

The Omnimodel and Modality Convergence

  • Explanation: True multimodal omnimodels are replacing fragmented single-modality models for text, audio, video, and code.
  • Evidence: Omni models handling text, audio, and video inputs natively without intermediate routing.
  • Practical implication: Developers no longer need to chain disparate models together for cross-modal tasks.

Startup Strategy and defensibility

  • Explanation: Startups must leverage focus and speed to build deep domain workflows before foundational models absorb their features.
  • Evidence: Historical tech ecosystem dynamics where platform shifts enable nimble startups to outpace incumbents in niche verticals.
  • Practical implication: Focus is the superpower of startups; founders must pick domains where they have deep proprietary insights.

Frameworks, Models & Processes

The Agent Harness Framework

  • How it works: Connects core LLM intelligence to execution environments, allowing models to take real-world actions via APIs, CLI, and SDKs.
  • Components:
    • Core model intelligence
    • Execution environment
    • Tool calling layer
    • Feedback loop
  • When to use: When designing autonomous developer tools, coding assistants, and multi-step agent workflows.

Examples & Case Studies

Logan Kilpatrick used AI Studio to build a gardening app on Android in minutes.

  • Illustrates: The speed and capability of AI-assisted native app development using Google developer tools.
  • Lesson: AI lowers the barrier to building functional software across platforms dramatically.

Actionable Takeaways

  • Immediate:
    • Leverage Google AI Studio for rapid mobile and web prototyping.
    • Adopt agentic workflows for coding and operational tasks.
  • Strategic:
    • Build applications that transcend simple prompt-response wrappers to avoid being 'eaten by the model'.
    • Focus on deep domain expertise and proprietary data loops.
  • Questions to investigate:
    • How fast will foundational models absorb current agentic frameworks?
    • What is the optimal balance between general-purpose models and vertical fine-tuning?

Claims Worth Verifying

  • Over 350,000 Android apps have been built using AI Studio. (metric)
  • Coding is the most robust general-purpose agent harness. (architectural theory)

Notable Quotes

"Everyone is going to be able to vibe code video games by the end of 2025." "We cannot let other people make the world a better place more than we can."

Compressed Summary

  • Agentic AI is transforming software development and product ecosystems.
  • Foundational models continually absorb external scaffolding and harnesses.
  • Google AI Studio and Gemini API provide robust infrastructure for developers and builders.
  • Startups succeed through extreme focus and deep domain execution rather than shallow wrappers.
  • Keywords: agentic, gemini, harness, developers, multimodal
  • Core insight: As foundational AI models advance and eat their own harnesses, developers and startups must build deep, domain-specific workflows rather than relying on transient wrapper logic.

Core insights

6
Mental Modelhigh noveltymoderate evidence

Advancing model capability absorbs external scaffolding: harnesses built around a model as a separate layer get internalized natively over time, so the durable abstraction boundary keeps migrating from application scaffolding into model weights.

Why it matters

It reframes the make/buy/harness decision: any custom orchestration layer is a depreciating asset whose value decays as a function of upstream model capability, which changes how you budget engineering effort and where you place proprietary value.

Generalization

For any layer that compensates for a capability gap in a fast-improving dependency, expect that layer to be commoditized or eliminated rather than extended.

External scaffolding is frequently absorbed by advancing model capabilities over time.
Open source video
Developers building custom harnesses risk being eaten by the model as intelligence improves.
Open source video
True differentiation moves upstream as raw model capability increases.
Open source video
Empirical Resultmedium noveltyweak evidence

Coding is claimed to be the most robust general-purpose agent harness — the coding loop (write, execute, observe errors, fix) is the substrate that generalizes to other autonomous tasks.

Why it matters

If general agency routes through the code-execution feedback loop rather than bespoke tool APIs, then investing in a strong code execution/sandbox/verification loop is a higher-leverage investment than building many narrow tool-calling integrations.

Generalization

Loops with executable, unambiguous feedback signals generalize better than loops relying on soft or human-judged feedback.

Coding has proven to be the most robust general-purpose agent harness.
Open source video
Architecturemedium noveltymoderate evidence

True multimodal 'omni' models consume text, audio, and video natively in a single model, removing the intermediate routing layer that multi-model pipelines require for cross-modal tasks.

Why it matters

Cross-modal pipelines lose an entire class of failure modes and latency/cost overhead (router errors, format conversion, context handoff) when modality routing is internalized into one model; the orchestration responsibility moves from the developer to the model.

Generalization

Whenever a routing or translation layer exists only to bridge incompatible components, expect consolidation into a single component with a broader native interface.

Omni models handling text, audio, and video inputs natively without intermediate routing.
Open source video
Developers no longer need to chain disparate models together for cross-modal tasks.
Open source video
Architecturelow noveltymoderate evidence

The agent harness is decomposed into four components: core model intelligence, execution environment, tool-calling layer, and feedback loop — establishing where each responsibility lives in an agentic stack.

Why it matters

Gives a concrete seam map for separating what the model owns from what the runtime owns, which is directly useful for deciding what to build, what to abstract, and what to expect the model to absorb next.

Generalization

Agent stacks can be evaluated by which component receives a given failure or capability — model, environment, tool interface, or feedback path.

Connects core LLM intelligence to execution environments, allowing models to take real-world actions via APIs, CLI, and SDKs.
Open source video
Practicemedium noveltymoderate evidence

Product evaluation criteria shift from engagement duration ('eyeballs') to task completion and user autonomy, meaning agent success metrics should be outcome-based rather than interaction-based.

Why it matters

Directly changes what you instrument: time-on-task and session length become suspect metrics, while task success, completion rate, and autonomy level become primary, which reshapes eval design and agent acceptance criteria.

Generalization

For autonomous systems, engagement-proxy metrics can reward failure modes that keep the user in the loop; outcome metrics require task-grounded ground truth.

Product metrics must focus on user success and autonomy rather than mere engagement duration.
Open source video
Tradeoffmedium noveltymoderate evidence

Defensibility for builders shifts away from wrapper logic toward deep domain workflows and proprietary data loops, with focus framed as the startup's structural advantage over incumbents.

Why it matters

It gives an explicit counter-position to 'build the harness' strategy: instead of betting on orchestration that the model may absorb, bet on domain depth and data feedback loops that the model cannot reach.

Generalization

When a general-purpose capability commoditizes a layer, durable advantage accrues to whoever owns the private data or workflow specifics that general models cannot synthesize.

Focus is the superpower of startups; founders must pick domains where they have deep proprietary insights.
Open source video
Build applications that transcend simple prompt-response wrappers
Open source video

Deep dives

5

Harness absorption rate and measurement

Research question

At what rate do foundational models internalize agent harness capabilities, and how much custom orchestration can be deleted per model generation while preserving task performance?

Why

If absorption rate is measurable, teams can set expiry budgets for scaffolding and shift investment to domain workflows and execution environments before their harness becomes stranded.

External scaffolding is frequently absorbed by advancing model capabilities over time.
Open source video
Developers building custom harnesses risk being eaten by the model as intelligence improves.
Open source video
Source video

Coding as general-purpose harness vs executable feedback

Research question

Is coding's robustness as an agent harness caused by its executable feedback loop, and does it outperform direct tool-calling on matched multi-step tasks?

Why

If the coding loop generalizes because of executable feedback, engineering effort should prioritize sandboxing, verification, and error recovery over building many narrow tool integrations.

Coding has proven to be the most robust general-purpose agent harness.
Open source video
Connects core LLM intelligence to execution environments, allowing models to take real-world actions via APIs, CLI, and SDKs.
Open source video
Source video

Outcome-based agent metrics and autonomy ground truth

Research question

Which task-completion and autonomy-level metrics predict agent product success better than engagement duration, and how can success predicates be instrumented reliably?

Why

Agent products that optimize engagement can reward keeping the user in the loop; outcome metrics require task-grounded ground truth, which changes eval design and acceptance criteria.

Product metrics must focus on user success and autonomy rather than mere engagement duration.
Open source video
Source video

Omni-model modality routing absorption

Research question

Do native omni-models match or exceed chained-modality pipelines on cross-modal accuracy, latency, and cost while removing the router layer?

Why

If omni-models absorb modality routing, cross-modal pipelines lose an entire class of failure modes and handoff overhead, but gain single-vendor dependency; architecture decisions should account for both.

Omni models handling text, audio, and video inputs natively without intermediate routing.
Open source video
Developers no longer need to chain disparate models together for cross-modal tasks.
Open source video
Source video

Replaceable harness seams

Research question

Which abstraction seams between model, execution environment, tool layer, and feedback loop minimize refactoring cost when a model natively absorbs one component?

Why

A seamed harness can swap components independently, so teams can adopt model-native capabilities without rewriting the whole agent and can isolate which component caused a behavior change.

Connects core LLM intelligence to execution environments, allowing models to take real-world actions via APIs, CLI, and SDKs.
Open source video
External scaffolding is frequently absorbed by advancing model capabilities over time.
Open source video
Source video

Article ideas

4

The Model Eats the Harness: Why Your Agent Framework Is a Depreciating Asset

Custom orchestration layers are depreciating assets: they should be built thin and replaceable, while durable investment belongs in domain workflows, proprietary data loops, and execution/feedback components the model does not yet own.

Angle

Treat harness code like inventory with a model-capability-driven write-down schedule; show what to build versus what to rent.

Source video

Code Execution Is the Harness That Survives Absorption

Coding is the most robust general-purpose agent harness because its execution environment and feedback loop provide executable ground truth, so builders should invest in sandbox/verification loops rather than broad tool-calling wrappers.

Angle

Compare coding-mediated agency to direct tool-calling and explain why feedback signal quality, not tool count, drives generalizable autonomy.

Source video

Stop Measuring Time-on-Task: Outcome Metrics for Agentic Products

Engagement duration is a misleading metric for agentic products because it can reward failure modes that keep users in the loop; task completion and autonomy level require task-grounded ground truth and should become primary.

Angle

Translate the metric shift into instrumentation requirements: success predicates, completion detection, autonomy levels, and validation against user-reported success.

Source video

Omni-Models Are the End of the Modality Router

Native multimodal omni-models internalize the routing layer that multi-model pipelines need, removing router errors and handoff latency but increasing single-provider dependency, so architectures should treat modality routing as replaceable rather than strategic.

Angle

Analyze the before/after pipeline, the failure modes removed, and the vendor substitutability risk that replaces them.

Source video

Project ideas

4

Harness Absorption Tracker

beyond-evals

For a fixed agentic task suite, the amount of custom orchestration code required to match a baseline success rate decreases measurably across successive model generations.

Proof of concept

Build a task suite with a modular harness; run it on multiple model versions; incrementally delete or disable harness components while measuring success until performance degrades.

Measurement

Percentage of harness components removed per model generation at iso-accuracy, plus LOC and latency changes.

Source video

Code-Loop vs Tool-Calling Harness Benchmark

movement-lab

A coding-mediated agent that writes and executes code achieves higher multi-step task completion than a direct tool-calling agent under matched tool availability.

Proof of concept

Implement two agents for the same tasks: one uses code execution to call APIs/CLI; the other uses function-calling tools directly; compare completion and error recovery.

Measurement

Task completion rate, steps to success, error recovery rate, and cost per completed task.

Source video

Outcome-Grounded Agent Eval Harness

beyond-evals

Task-completion and autonomy-level metrics correlate with downstream user retention better than engagement duration for agentic products.

Proof of concept

Instrument an agent product with explicit success predicates and autonomy levels; log engagement duration; validate both against user-reported success and retention.

Measurement

Correlation coefficients between outcome metrics vs engagement metrics and retention/user-reported success.

Source video

Replaceable Harness Seam Kit

new

Enforcing hard abstraction seams between model, execution environment, tool layer, and feedback loop reduces refactoring effort when a model natively absorbs one component compared with a monolithic agent.

Proof of concept

Build the same agent twice (monolithic vs seamed); replace the tool-calling layer with a model-native capability; measure code changes, regression test pass rate, and time to swap.

Measurement

LOC changed, files touched, time to integrate, regression test pass rate, and performance delta.

Source video

Architectural implications

4

Capability responsibility migrates from developer-authored scaffolding into the model itself.

Before

Application code owns orchestration, prompt shaping, and task-specific wrappers around a general model.

After

The model natively performs those behaviors, leaving the application layer with domain logic, data, and distribution.

Consequence

Custom harness code becomes technical debt at an unpredictable rate; architecture should favor thin, replaceable orchestration and avoid deep investment in capability-substituting layers.

Source video

Modality handling consolidates from multi-model pipelines into a single omni-model call.

Before

Text, audio, and video are handled by separate models joined by a routing/translation layer.

After

One model ingests multiple modalities natively without intermediate routing.

Consequence

Removes router-level failure modes and handoff latency but increases dependence on a single vendor model's interface and availability.

Source video

Agent harness decomposed into model intelligence, execution environment, tool layer, and feedback loop.

Before

Agent frameworks blend model calls, tool wiring, and orchestration into one bespoke codebase.

After

Responsibilities are separable, so each can be replaced as models absorb adjacent capabilities.

Consequence

Enables targeted upgrades (swap feedback loop or execution env) without rewriting the whole agent, and clarifies which component to blame when behavior degrades.

Source video

Evaluation shifts from engagement-based to task-outcome-based measurement.

Before

Success measured by session length, impressions, or interaction volume.

After

Success measured by task completion and user autonomy.

Consequence

Requires task-grounded ground truth and completion detection, which is harder to instrument than engagement telemetry.

Source video

Tradeoffs and failure modes

5

Building custom agent harnesses

Benefit

Near-term control, task-specific performance, and differentiation before native model support exists.

Cost or risk

The harness may be absorbed by model improvements, converting the investment into stranded scaffolding with shrinking defensibility.

Developers building custom harnesses risk being eaten by the model as intelligence improves.
Open source video
Source video

Wrapper-layer products

Benefit

Fast to ship on top of a general model with minimal domain investment.

Cost or risk

Simple prompt-response wrappers are the most exposed to being internalized by the model, leaving no durable moat.

Build applications that transcend simple prompt-response wrappers
Open source video
Source video

Single omni-model for cross-modal work

Benefit

No chaining of disparate models; native multi-modality without intermediate routing.

Cost or risk

Consolidates cross-modal dependencies onto one model interface and provider, reducing substitutability.

Developers no longer need to chain disparate models together for cross-modal tasks.
Open source video
Source video

Optimizing for engagement metrics

Benefit

Familiar, easy-to-instrument telemetry (time-on-screen, impressions).

Cost or risk

Rewards keeping users in the loop rather than completing tasks, misaligning agent design with user autonomy.

Product metrics must focus on user success and autonomy rather than mere engagement duration.
Open source video
Source video

Startup breadth versus focus

Benefit

Extreme focus enables speed and depth in a chosen vertical before incumbents react.

Cost or risk

Narrow domain bets fail if the domain is not one where the model cannot absorb the workflow.

Focus is the superpower of startups; founders must pick domains where they have deep proprietary insights.
Open source video
Source video

Open questions

5

How fast will foundational models absorb current agentic frameworks?

Why unresolved

The summary poses this as an open question without providing an absorption rate, benchmark, or timeline evidence.

Research direction

Track regression of harness-specific capabilities over successive model releases: measure how much orchestration code can be deleted per model generation while preserving task performance.

Source video

What is the optimal balance between general-purpose models and vertical fine-tuning?

Why unresolved

Stated as an unanswered question; no cost, accuracy, or maintenance data is offered.

Research direction

Benchmark domain tasks across general models, prompting/tooling, and fine-tuned variants on accuracy, latency, cost, and update overhead.

Source video

Is coding actually the most robust general-purpose agent harness, or is that an artifact of executable feedback availability?

Why unresolved

The claim is asserted as an architectural theory without comparative evidence against non-coding harnesses.

Research direction

Run matched multi-step agent tasks with coding-mediated execution versus direct tool-calling and compare completion rates and error recovery.

Source video

How do you reliably measure task completion and autonomy in place of engagement metrics?

Why unresolved

The summary asserts the metric shift but provides no measurement mechanism or ground-truth definition.

Research direction

Build outcome-labeled evals with explicit success predicates and autonomy levels, and validate them against user-reported task success.

Source video

Which parts of an agent stack should be kept thin to remain replaceable as models absorb capability?

Why unresolved

The absorption principle is stated but not translated into concrete ownership boundaries or interfaces.

Research direction

Prototype harnesses with hard abstraction seams (model, environment, tools, feedback) and measure refactoring cost when a model natively absorbs one component.

Source video

Key claims

8
factualVerification needed

Antigravity is the agent harness powering Google products and serves as a unified agentic layer across search, cloud, and AI Studio.

Evidence

Antigravity is the agent harness powering Google products.

Question

Which Google products run on Antigravity, and is it exposed or documented for external developers?

Source video
causalVerification needed

External scaffolding is absorbed by advancing model capabilities over time.

Evidence

External scaffolding is frequently absorbed by advancing model capabilities over time.

Question

What concrete harness capabilities have been internalized across model generations, and at what rate?

Source video
comparativeVerification needed

Coding is the most robust general-purpose agent harness.

Evidence

Coding has proven to be the most robust general-purpose agent harness.

Question

How does a coding harness compare to non-coding tool-calling harnesses on matched multi-step tasks?

Source video
factualVerification needed

Over 350,000 Android apps have been built using AI Studio.

Evidence

350,000+ Android apps built using AI Studio.

Question

Is the 350,000 figure unique apps, projects, or builds, and over what time window?

Source video
causalVerification needed

Omni models handle text, audio, and video inputs natively without intermediate routing, removing the need to chain models for cross-modal tasks.

Evidence

Omni models handling text, audio, and video inputs natively without intermediate routing.

Question

Do omni models match or exceed chained-modality pipelines on cross-modal accuracy, latency, and cost?

Source video
opinionVerification needed

Product metrics should shift from engagement duration to user success and autonomy.

Evidence

Product metrics must focus on user success and autonomy rather than mere engagement duration.

Question

Which task-completion metrics predict downstream retention better than engagement duration for agentic products?

Source video
predictionVerification needed

Everyone will be able to vibe code video games by the end of 2025.

Evidence

Everyone is going to be able to vibe code video games by the end of 2025.

Question

What level of game complexity is achievable end-to-end by non-programmers with current agentic tooling?

Source video
causalVerification needed

True differentiation moves upstream as raw model capability increases.

Evidence

True differentiation moves upstream as raw model capability increases.

Question

Which startups retained defensibility after adjacent model capabilities were absorbed, and what did they own?

Source video

Connections

5