Koray Kavukcuoglu on frontier models, coding agents, and building AGI

A frontier race without a finish-line test

Koray Kavukcuoglu’s message this week is deliberately uncomfortable for anyone looking for a clean AGI scoreboard: Google DeepMind does not appear to believe that one definitive test will settle the question of artificial general intelligence. In the September 1 material around his Google for Developers appearance, the central point is that there is no accepted threshold or benchmark that can determine when AGI has arrived [1]. That is not a small caveat. It changes the framing from “which model passed the exam?” to “which systems are becoming broadly useful, reliable and trusted across more of the work people actually do?”

The distinction matters because the public AI debate still tends to chase leaderboards. A model beats another model on code, math, multimodal reasoning or tool use; a lab declares progress; rivals respond with a stronger score. Kavukcuoglu’s argument, as summarized in the fresh reporting, is that AI progress has historically moved through shifting targets rather than through a single canonical evaluation [1]. In this view, AGI is not a trophy awarded by one benchmark suite. It is a compounding process in which systems become capable enough, dependable enough and context-aware enough to be used across widening domains.

That does not mean Google DeepMind is stepping away from the race. Quite the opposite. The Decoder reported on September 1 that Kavukcuoglu pushed back sharply against the idea that Google is content to compete mainly on distribution or price-performance rather than raw model leadership [2]. He said, in effect, that being at the frontier is the strategic priority, while also acknowledging that Google’s current models are “a little bit below the frontier” [2]. That combination — candor about today’s gap, insistence on tomorrow’s ambition — is the most revealing part of the story.

Gemini 4: ambition, but not yet proof

The near-term symbol of that ambition is Gemini 4. Kavukcuoglu described the Gemini 4 effort as Google DeepMind’s most ambitious pre-training run so far, and said it was progressing well, while still emphasizing caution before the model proves itself internally and externally [1]. The Decoder similarly noted that he offered no concrete new frontier result and no firm detail beyond the characterization of Gemini 4 as the “most ambitious run” so far [2].

That restraint is important. In a market accustomed to launch theatrics, Google DeepMind is signaling both scale and uncertainty. A large pre-training run can consume vast compute and organizational focus, but it is not a product until it survives post-training, evaluation, safety work, latency constraints, integration and user feedback. Kavukcuoglu’s framing suggests that Google does not want Gemini 4 judged only by the size of the run. The stronger test will be whether it produces systems people can use, trust and delegate to.

There was also no meaningful update on Gemini 3.5 Pro in the fresh reporting, with The Decoder noting that the model was still apparently in the works and months late [2]. That absence leaves a strategic tension. Google DeepMind wants to convince developers and enterprise customers that it will own the frontier, but the evidence for that claim remains partly future-dated. As AI Daily Post put it on September 1, the rhetoric is confident while the evidence is still pending [4].

Why coding agents sit at the center

The most concrete technical theme in Kavukcuoglu’s remarks is not a new benchmark claim but the shift from language models to agents. The recent Gemini story, especially across the Flash line, is being presented as a move toward systems that can use tools, make decisions in workflows and execute software tasks rather than merely generate text [1]. The Decoder reported that Kavukcuoglu described the progression from Gemini 3.5 through 3.6 to 3.7 as a transformation from model to coding agent [2].

That emphasis is not accidental. Software engineering is a unusually rich environment for measuring whether an AI system is becoming more general. Code requires planning, tool use, debugging, memory of constraints, interaction with repositories, tests, terminals and sometimes other agents. It also gives relatively sharp feedback: the program runs or it does not; the test passes or fails; the bug is fixed or remains. Kavukcuoglu called software engineering the most critical domain and environment for these systems to succeed in [2].

The Flash focus is particularly notable because “Flash” historically implied speed and cost efficiency more than maximum intelligence. In the September 1 summaries, however, the Flash series is portrayed as the track where Google has been iterating fastest toward agentic performance [1]. That has strategic implications. If Google can turn faster, cheaper models into capable agents, it may compete not only on benchmark peaks but on scalable deployment — agents that can be run frequently, embedded deeply and improved through real user workflows.

AGI as deployment plus feedback

The deeper thesis is that AGI is built with users, not merely announced from a lab. The September 1 8news summary says Google DeepMind is treating the path to more general AI as co-development with users, where everyday use cases such as email assistance sit alongside scientific research applications and help identify which capabilities matter in practice [1]. That is a classic Google advantage if it works: massive distribution, product telemetry, infrastructure and research talent feeding the same loop.

This approach also explains why Kavukcuoglu talks about trust. A model that scores well but cannot be relied on in an ambiguous workflow remains a demo. An agent that can collaborate, recover from errors, ask for clarification, use tools safely and improve through feedback becomes infrastructure. The difference is especially important for coding agents, where autonomy without reliability can create security bugs, broken builds or hidden technical debt.

The same-day Google developer material on agentic video understanding shows how broadly the “agentic” framing is being applied inside the Gemini ecosystem. In that example, Gemini analyzes video by selectively calling tools such as transcript, frame and audio retrieval instead of ingesting an entire video at once [5]. Although that is a separate product capability, it reinforces the architectural direction relevant to Kavukcuoglu’s remarks: Gemini is being pushed toward systems that decide what information they need, gather it through tools, observe the result and continue iterating [5].

The benchmark problem is also a trust problem

Saying that AGI has no definitive benchmark is not the same as saying benchmarks do not matter. They still discipline claims, expose weaknesses and let customers compare systems. But Kavukcuoglu’s position implies that benchmarks are directional signals, not final authority [1]. That is especially true when agents are involved, because the performance of an agent depends on model capability, tool design, scaffolding, memory, permissions, latency and the surrounding product.

This is why the admission that Google is currently somewhat below the frontier is striking. It narrows the debate to execution. Google DeepMind says it has the team, resources and full technical stack to close the gap [2]. HelloMarvis’ September 2 summary interpreted the comments as a strategic bet on long-term capability rather than short-term efficiency [3]. The open question is whether that full-stack advantage translates into models and agents that users experience as clearly better.

For developers, the practical takeaway is cautious optimism. Google DeepMind is not pretending that a single AGI test will crown a winner. It is arguing that frontier leadership will emerge through systems that can do more, act more reliably and earn more delegation from users. But it is also asking the market to wait for Gemini 4, for stronger Pro updates and for Flash’s agentic gains to become undeniable in production.

What to watch next

The next phase of this story will not be decided by slogans about AGI. It will be decided by release quality, developer adoption, independent evaluations and whether Gemini agents can perform long-horizon work with fewer failures. Watch for three signals: a concrete Gemini 4 reveal, clearer evidence that Flash models can handle complex coding workflows, and a more explicit account of how Google DeepMind measures trust in agentic systems.

Kavukcuoglu’s core claim is that Google must be at the frontier and will get there [2]. His more subtle claim is that the frontier itself is changing. It is no longer just the model with the highest score on a static test. It is the system that can reason, use tools, cooperate with people, survive real workflows and keep improving from the feedback loop. That is a harder race to judge, but it is probably closer to what AGI will actually feel like when it arrives.

Sources from the last 72 hours

1. [1]Koray Kavukcuoglu on frontier models, coding agents, and building AGI · Google · 8news.aiSep 1, 2026, 3:00 PM UTC
2. [2]Google Deepmind's new chief says frontier AI leadership is the only thing that mattersSep 1, 2026, 12:00 AM UTC
3. [3]DeepMind chief: frontier AI leadership is all that mattersSep 1, 2026, 7:01 PM UTC
4. [4]DeepMind Chief Vows to Lead Frontier AI RaceSep 1, 2026, 12:00 AM UTC
5. [5]Agentic video understanding in Gemini · Google · 8news.aiSep 1, 2026, 5:00 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.