Kill It or Keep It

Twenty minutes into his conversation with Bloomberg's Ed Ludlow, Mike Krieger had to explain a piece of news that was, by his own count, about two hours old. Anthropic had just released Mythos — a new model, deliberately withheld from the general public and handed instead to a small group of cybersecurity partners and open-source maintainers. Krieger co-leads Anthropic Labs. He hadn't planned to make news on the Kaleidoscope stage. He just happened to be the person in the room when the model shipped.

That's the version of Labs that came through all session: less a research division, more a discipline for deciding what's real before the rest of the company has to bet on it.

The rule that keeps Labs honest

Krieger, Instagram's co-founder and Anthropic's former CPO, described Labs as an internal incubator — not a lab in the traditional sense, but a small team explicitly excused from the roadmap. "If it would have belonged on another team's roadmap, it doesn't belong in Labs," he said. The point isn't to duplicate work other teams are already prioritizing. It's to sit somewhere nobody else has bandwidth to look: what isn't possible with the current models, but might be if the next one lands.

That freedom comes with a governor. Every two weeks, Labs reviews every project and makes a binary call: keep going or kill it. "The opportunity cost of working an extra even two weeks on the wrong project is pretty high," Krieger said, given how stacked the team is with talent and how fast the models move underneath them. There's no room for the nine-month pet project that quietly stops producing.

Two things came out of the first version of Labs that make the case for the model: Claude Code and the Model Context Protocol. Both graduated into full teams — MCP into an open-source standard now used well beyond Anthropic. But Krieger was just as interested in what didn't graduate. An early computer-use prototype could technically click around a desktop, but it solved problems in ways too idiosyncratic to trust. Rather than kill the idea outright, the team kept the half-broken prototype on the shelf and re-ran it against every new model. It failed for months. Then, on Claude 3.7, it didn't. "The whole goal is that you're showing as much about what's not possible yet with the model, but might be, as they get better than it is about doing it today."

Testing on 2,000 people before it ever reaches the outside

Internally, Anthropic calls its dogfooding practice "antfooding" — a nod to the company's self-deprecating "we're Anthropic, we're ants" culture that Krieger admits he didn't love at first. What's changed is the scale of it. When he joined, the internal and external versions of Claude were nearly identical — a sign, in his read, that the team wasn't being bold enough. Now internal builds ship and unship constantly, tested against a population of more than 2,000 employees before anything reaches a customer.

Cloud Code's internal test population is high-volume enough, and high-signal enough, that Krieger compared it to an on-call rotation: someone is always watching for what breaks. That internal feedback loop is also, in his telling, a meaningful piece of why the infrastructure underneath Cloud Code became the foundation for Anthropic's other products. Co-Work runs on the same base. Founders building their own agents have told him they scrapped their own frameworks entirely once they got access to the underlying agent SDK.

Why the newest model isn't going to the public first

Mythos is where the conversation stopped being retrospective. Krieger was direct about what's actually happening: a model that vastly outperforms Opus 4.6 on coding and agentic tasks, deliberately not released broadly, because the same capabilities that make it exceptional at general-purpose engineering also make it exceptional at offensive cybersecurity. "You can't isolate a capability," he said. Train a model to be excellent at debugging and writing code, and you've also trained something that's excellent at finding and exploiting vulnerabilities — whether or not that was the goal.

The response wasn't to sit on the model. It was to get it into the hands of the people building defenses first — cybersecurity partners and open-source maintainers — before wider distribution, so the industry has lead time to build inoculation rather than discovering the gap after the fact. "This is the worst the models will ever be," Krieger said, describing a mantra the Labs team repeats to itself. Capability only goes up from here, unevenly across domains but consistently over time, which means the safeguards have to be built for where the model is heading, not just where it is now.

He also pointed to a shift already underway inside Anthropic: moving away from what he called "approval theater," where a human signs off on every model action until they stop actually reading what they're approving. The alternative Anthropic has been testing internally — and has since shipped publicly — has one model evaluate what another is doing and flag what needs a human versus what's safe to proceed. It's a small architectural choice with a large implication: the safeguard has to scale with the model, not with the patience of the human watching it.

The constraint underneath everything is compute

None of this happens without capacity, and Krieger didn't dodge how tight that constraint still is. Anthropic was an early multi-chip bet — GPUs, TPUs, Trainium, Trainium 2 — run for years by what he described as a small inference team "multiplexing their attention" across all of them. That bet is paying off now partly because the models themselves have gotten good enough to help write the infrastructure code that supports them, freeing up a team that's since grown substantially.

Krieger's own workaround for the constraint is Cloud Overnight — queuing a full night's worth of work before bed specifically because fewer engineers compete for that capacity window. "It's actually a funny ritual to wake up in the morning and have to go, oh, this thing — good morning," he said. It's a small, personal example of a company-wide problem: demand for compute is outpacing even aggressive capacity investment, including the multi-gigawatt TPU commitments and partnerships announced the same week as this conversation.

What Krieger actually wants to build

Asked for a preview of what hasn't been built yet, Krieger didn't reach for a product name. He reached for a distinction. Right now, long-horizon autonomy — a model working productively for 30 hours unsupervised, the way some early enterprise customers have — is the exception, not something most people can access. His stated focus for what comes next is closing that gap: models that function as a productive coworker over long stretches of unsupervised work, not just a fast delegate you check in on every few minutes.

It's a smaller claim than "the best AI products haven't been built yet" sounds like on a session title slide. But it's the more useful one. The products Labs is chasing aren't defined by a new capability landing all at once. They're defined by whether a two-week review, an antfooding population of 2,000, and a staged release to the people building defenses first can keep pace with a model that gets less predictable every time it improves.

Watch the full session on-demand.

Keep the conversation going. Register now for HumanX 2026 Amsterdam (September 22–24, 2026) or HumanX 2027 Las Vegas (March 7–10, 2027). Be sure to subscribe to The Output newsletter and save yourself from bad AI takes (and save on your HumanX ticket).