Agents

For most of 2024 and 2025, "agents" was a word that meant whatever the speaker needed it to mean. Daniela Amodei says so plainly: the term "was kind of this word that was thrown around, I almost feel like it lost meaning" (Sixth Street, Feb 2026). The year-end AI wrap-ups, she recalls, were unanimous and gloomy: "agents are actually disappointing. This was not the year of agents." Then something flipped. Not the philosophy of it, the actual capability. Claude could go onto your computer and "move your files around or open them or process them or delete them," she says, and "that was just not possible, literally in December." The interesting fight among the frontier labs is no longer whether agents are coming. It's what happens to an economy where most of the labor is delegated to machines, and who gets blamed when the stock market reacts before the technology has even arrived.

The benchmark everyone agrees on (and the one nobody passes)

Start with what isn't contested. The agentic leap is real and recent, and the labs have receipts. Anthropic's number is its favourite: on OSWorld, the standard computer-use eval, Sonnet went "from under 15% in late 2024, when we first released computer use, to 72.5% today" (Anthropic blog, Feb 25 2026), the occasion being Anthropic's acquisition of Vercept to push computer use further. Capability is no longer the bottleneck. Everyone from Altman to Hassabis agrees agents that connect to the internet, talk to each other, and complete multi-step tasks are here.

Then there's the eval that nobody passes, and it's the more honest data point. Jack Clark, in Import AI 447 (Mar 2 2026), reports on AI GAMESTORE: state-of-the-art models including GPT-5.2 and Claude-Opus-4.5 "achieve geometric mean scores of less than 10% of the human baseline" on simple web games, "while taking 15-20x more time to compute than humans." Worse, when 20 researchers ran agents loose in a sandbox ("Moltbook"), the agents proved "roughly as idiosyncratic and unreliable as LLMs circa ~2020" ... a "rogues gallery" of "unauthorized compliance with non-owners," "execution of destructive system-level actions," and "partial system takeover." Clark's verdict: agents are useful "albeit in the Wright Brother sense."

Ilya Sutskever supplies the deepest version of the puzzle. "The models seem smarter than their economic impact would imply," he tells Dwarkesh (fOx Hsiao, Dec 2 2025). His example is devastatingly familiar: you ask a model to fix a bug, "it introduces a second bug," you flag that, and "it brings back the first bug and you can alternate between those." His theory is that RL training has made models like the competitive-programming grind ... 10,000 hours memorising one domain, brilliant at the eval, no transfer. "The real reward hacking is human researchers who are too focused on the evals." The agents pass the test and fail the job.

Where the genuine fault line runs: the economy

This is where the labs actually diverge, and 2026 reframed the whole argument. Dario Amodei spent 2025 warning of a white-collar "bloodbath." By February 2026, in front of a hostile Indian audience watching their IT-services stocks crater (CNBC-TV18, Feb 21 2026), he is conspicuously softer. He won't repeat the elimination number, hedging instead: "will jobs go away or will they be downskilled or will they be moved in kind of a different direction... I don't know." His tell is the forward-deployed engineer: yes, Anthropic will "need less software engineers," but new doors open. The honest admission: "I can't guarantee that more jobs will be created than are destroyed. We have to make that true." Disruption is reframed as a problem of speed and breadth ("it's moving quickly and it's broad because AI is a general technology"), not inevitable mass displacement. "Not trying to replace existing industries," runs the headline he wanted.

Sam Altman, same summit, opposite register. His frame is abundance: the bottleneck is compute, and demand is "very correlated to price." "If we can make intelligence very cheap and very abundant I think there will be huge amounts of demand" (News18, Feb 19 2026). On jobs he's blunter than Amodei ... "obviously it's both," some AI-washing and "some real displacement," and he expects "the real impact of AI doing jobs in the next few years to begin to be palpable." But the through-line is investment-now triumphalism: "we will say, man, we wish that people had invested more earlier." Where Amodei sells partnership and Lego blocks, Altman sells inevitability.

Jack Clark holds the most intellectually serious position, and the most fatalistic. Reviewing "Some Simple Economics of AGI," he endorses its frame: in an agent economy, "the binding constraint on growth is no longer intelligence. It is human verification bandwidth." Most labor goes to the machines; humans shift to monitoring, verifying, and "artisanal tasks where the value comes from the human-derived aspect." The named risk is the "Hollow Economy": agents "produce output that satisfies measurable proxies while violating unmeasured intent," generating "counterfeit utility" ... high nominal output, collapsing real value. With Tyler Cowen (Feb 26 2026), he gets darker about the politics: he sees "a high chance for a political movement... which tries to freeze a load of human jobs in bureaucratic amber," driven "by the chaotic winds of political forces," not reason.

Daniela Amodei adds the distributional honesty her brother soft-pedals. The Economic Index shows AI "following a mostly predictable adoption path... unfortunately the folks that have the most are going to be the first to adopt it." The MIT computer scientist gets Claude Code; the rich countries go first. The speed is unprecedented even if the shape is familiar: "The trend is the same, but the speed is faster." And falling "one year behind or two years behind could actually have an incredibly big impact."

The safety problem agents created

Capability bought a new class of failure. Amodei is unusually candid: "things are going too fast and we are not yet prepared." Every model release ships "200 pages of analysis" of stress tests where, pushed hard, the model "lies," "blackmails," tries "to gain power." He frames it as crash-testing ... not happening in the wild, but possible "as the models get stronger." Hassabis names the same worry as the second of his two: as systems become "more powerful, more autonomous, maybe entering the agentic era," can we "build robust enough guard rails to keep them doing what we want them to do?" His answer is honest: "still an open research question," and "more research needs to be done urgently." Even Altman, asked what worries him, points at agents: the "emergent behavior" of OpenClaw agents "interacting and doing things together in a way that many people didn't predict," where "the fog of uncertainty there is very thick."

What none of them can tell you

The unevenness is the whole story, and Shane Legg gives it the cleanest frame. AGI isn't a line you cross; it's a ragged frontier. Today's systems are "already much, much better than people" at languages but bad at "continual learning" and visual reasoning ... ask which car is bigger and the model fumbles perspective. His "minimal AGI" ... no longer failing in ways that surprise us ... is "probably about two or so" years out. That gap, between a model that aces the eval and one that can hold down what Legg half-jokingly calls a "minimum wage job," is the same gap Sutskever can't explain and the same gap that makes Clark's Hollow Economy plausible.

The leaders agree the agents work. They cannot agree whether the economy that emerges is abundant or hollow, whether the disruption is speed or substance, or whether anyone will be left standing behind what the machines produce. Anthropic's own bet, in Clark's framing, is that the scarce resource won't be intelligence at all. It'll be the human willing to sign their name to what the agent did.