Gemini Fell Off the AI Recommendation List. That's a Map of Where Value Moved.

Here is the most useful sentence written about AI competition this month, and it comes from an absence. In Simon Willison's read of Ethan Mollick's evolving guide to which AI to use, Gemini has quietly dropped off the list.

Not because Google's models got worse. Because the question changed. A year ago, Mollick's guide was a chat-model comparison: ChatGPT, Claude, Gemini, with o3, Claude 4 Opus, and Gemini 2.5 Pro as the headline models. You picked a chatbot and talked to it.

Today the guide is "much more about agentic systems", described as "where the AI is capable of doing the equivalent of many hours of real human work in one go." The categories are now Codex, ChatGPT Work, Cowork. And Gemini fell off, per Willison's summary, "since Google still doesn't have an established entry in the Codex/ChatGPT Work/Cowork category."

That is not a review. It is a category boundary being redrawn in public. When the same trusted guide swaps its entire axis of comparison inside twelve months, the interesting signal is not who scored highest. It's which axis got retired.

The chat-model beauty contest is over. The evaluators stopped asking "which model answers best" and started asking "which platform can be handed work." Google has a model that competes at the frontier and still gets dropped, because the thing being ranked is no longer the model.

This is a Wardley Mapping moment, and it's worth being precise about what moved on the map.

A year ago, the conversation model sat in the "product" stage of evolution: differentiated, compared spec-by-spec, the thing you chose. Today it's sliding toward commodity. The layer that now carries the differentiation is the harness around the model: the environment that lets an agent run for hours, hold tools, keep state, and come back with finished work.

That is the Harness Hypothesis in one sentence. The value in AI isn't in the model; it's in the harness that connects the model to the world. Mollick's guide is now a harness comparison wearing model names.

You can see the commoditization in the same week's news. Anthropic shipped Claude Opus 5 with official messaging that it "comes close" to Fable-level performance while its efficiency "only just matches GPT 5.6 Sol." Frontier models are converging and being compared on price. Meanwhile, Moonshot released the 1.56TB Kimi K3 open weights, a 2.8-trillion-parameter model you can just download.

When frontier capability leaks into open weights and rivals cluster within a few points of each other, the model stops being the moat. Whoever owns the harness that turns those interchangeable models into completed work owns the user relationship. That's Aggregation Theory: the platform that aggregates the demand for finished work, then commoditizes the model supply underneath, wins.

Walk the argument forward. If the model is commoditizing, the fight relocates to three places: the runtime that executes multi-step work, the tools the agent can reach, and the observability that makes long autonomous runs safe to trust. Every one of those is a harness concern, and the week's releases read like a scorecard of who is building there.

OpenAI's agents SDK shipped v0.14.0 with Programmatic Tool Calling, which lets supported models generate code that coordinates several tools at once and trims the intermediate results before they clog the run. That is not a model improvement. It's plumbing for agents that do many things in sequence without a human in the loop. The same release makes model and tool data stop being logged by default unless you opt in, and emits task and turn spans automatically so a long run is legible after the fact.

Read those two together. Coordination and observability. They are exactly the capabilities you need to hand an agent "many hours of real human work" and still sleep at night. This is the layer Mollick is now ranking, and OpenAI is shipping it as product feet-first.

Meanwhile, the picks-and-shovels vendors are converging on the same story from the neutral side. Pydantic's agent framework reached v2.19.0 with better handling of provider errors and cancellation partway through a run. Arize Phoenix shipped v19.9.0 with trace panels that let you expand or collapse all span info at once, because when your agent takes forty steps, reading the trace is the actual job. The tooling ecosystem has quietly reorganized itself around watching agents work, not around chatting with models.

Now put Google back in the frame. The company is, by one analyst's argument, "the closest strategic mirror of NVIDIA's full-stack method", owning chips, models, and infrastructure end to end. The airline metaphor in that piece is apt: the engine matters, but so does the airframe, the airports, the scheduling software, the route network. Google owns a superb engine. What Mollick's guide exposes is that it hasn't yet flown a route in the work-automation category. Gemini Spark, per Willison, "has yet" to establish an entry there.

This is the gap between capability and product placement. Being able to build the harness is not the same as having shipped one users can be handed work through. Commoditize Your Complement explains part of Google's discomfort here: it would love the model layer to stay the decisive layer, because that's the layer where its full-stack advantage compounds. But the market moved the decision one layer up, into the harness, and that's a layer defined by developer adoption, tool ecosystems, and enterprise workflow integration, not by TPU economics.

There's a security subtext too. Once agents run for hours coordinating tools with data logging off by default, the Trust Boundary Model gets harder, not easier. The relay market investigation into resold LLM tokens (pooled API keys, proxied through "unprotected support bots") is a reminder that every place work crosses a trust level is a place to abuse. The harness layer isn't just where value accrues. It's where the attack surface for ai agent security 2026 concentrates.

The obvious objection: one man's guide is not a market. Mollick is a single influential voice, and "Gemini fell off a recommendation list" is a long way from "Google lost the agent category." Google could ship a Codex-class product next quarter and reappear at the top. Fair.

And the guide is opinionated by name. It reflects one professor's workflow, weighted toward the categories he personally uses to get work done. Absence from it is not market-share data.

But that undersells why this particular signal matters. The point is not the ranking; it's the axis change. When a trusted evaluator retires "which chatbot" and replaces it with "which work platform" inside a year, that reflects a shift in what buyers are trying to buy, and evaluators track buyers. The signal survives even if Google reappears next month, because reappearing would mean Google shipped a harness, which concedes the entire premise: the model was never the thing being ranked.

Google's full-stack strength is real and may prove decisive over a longer horizon. The near-term reading stands regardless. Right now the category is defined by the harness, and Google is competing in it with a model.