Two weeks before Anthropic priced its $81 billion IPO, Labs co-lead Mike Krieger described on the BoxWorks stage how his team builds when software takes hours. The scarce resources he kept returning to sit outside the code.
IPO figures as reported by the Associated Press; first-day trading as reported by The Next Web and Forbes. Krieger's figures are his own estimates from the stage.
AAnthropic priced its initial public offering at $83 a share on November 18, above the $70 to $78 range it had marketed, and raised about $81 billion at a valuation near $1.5 trillion, according to the Associated Press. The shares opened on Nasdaq under the ticker ANT at $111.50 the next day and closed at about $97, roughly 17% above the offering price, on a day when the broader market fell. Two weeks before the pricing, at BoxWorks in San Francisco, Mike Krieger gave a working-level account of how the company behind that number builds its products.
Krieger co-founded Instagram and spent eight years there. He served as Anthropic's chief product officer until January, when he moved to co-lead Anthropic Labs, the group behind Claude Code, the Model Context Protocol and Cowork, according to eWeek. I was in the audience on November 5 when Box CEO Aaron Levie interviewed him on the main stage. He described an engineering organization where building has become cheap and the scarce resources have moved elsewhere.
Krieger compressed a six-month build into an afternoon
Krieger said the change feels normal week to week and enormous when you zoom out. He pointed to a chart of commits into Anthropic's main codebase from May to September, which Levie described as close to vertical. Anthropic published its own figures for that stretch in October: merges rose from about 2,500 a month in May to more than 30,000 a day in September, the load that pushed the company to replace Git internally.
Mobile changed what Instagram's team could build, he said, but never at a week-over-week pace like this one. Asked how he would build Instagram today, he said current models can produce a fully featured version in about four hours, compared with the six months of trial and error the original took. Levie summarized the ratio as six months to about three hours, and Krieger agreed. That figure is his estimate for a working prototype; running a service at Instagram's scale remains a separate job.
He has the model draw the interface before anyone writes code
With the build time gone, Krieger asked what should fill the months it used to occupy. His answer was customer contact: put a prototype in front of customers early and listen. In his own work, he has the model generate the user interface first, talks it through, and has Claude critique the design before any code exists. People find it harder to disagree with a finished product, he said, so a sketch along the way keeps the debate open while changes are still cheap.
Krieger runs 85% of his engineering from Slack
Asked what his team's working interface looks like, Krieger estimated that his own engineering now happens about 85% in Slack and 15% in the terminal. He described a common internal pattern. An employee posts a point of friction with one of Anthropic's products in a feedback channel. Claude responds first, tries to reproduce the problem, and often has a pull request open by the time a person reaches the thread.
For new products, he works in a channel organized into many threads. The agents behind it number in the dozens to hundreds, and he experiences the work as delegating to a collaborator. Teammates join the same channel and work with Claude together. Anthropic had to teach its models manners for that setting. Krieger said one theme across model generations has been the social dynamics of a group chat, including when to stay quiet. His teams have told Claude to jump in only when someone tags it.
For an IT leader, the design choice matters more than the percentage. Anthropic puts the agent in the channel where people already report problems. That also puts a code change one message away from anyone in the channel, so the approval path is the piece your security and change-management teams need to design before you copy the pattern.
Fred Brooks still applies when agents write the code
Krieger warned that cheap rewrites leave old failure modes in place. He cited second-system syndrome from Fred Brooks's The Mythical Man-Month: a team that rebuilds a system can forget what it learned the first time or repeat the same mistakes. He wants the edges of a product to stay open to continuous rewriting, and he sees value in keeping the architectural history that explains why the core looks the way it does.
When code costs almost nothing to regenerate, the design record and the reasons behind it become the asset to protect.
Ask who owns that record in your organization, and whether it would survive the next reorganization or platform switch.
Customers now set the speed limit
Levie pressed on a point every enterprise will recognize. Software can now be built in hours, but customers, rollouts and the real world move at their own pace, and engineers and product managers start to act as the interface to that world. Krieger agreed that putting a product in front of real users is still where teams find out what they got wrong.
He added a constraint specific to enterprise software. A vendor cannot sit down with every person at every company that rolls out its product, so the product has to help people get started. He pointed to Claude's Artifacts as one direction, where the interface can be a structured view or a diagram the user keeps editing, and to the idea of a companion next to every person who uses a product.
He also described how Labs is organized. Each product manager covers two or three early products, and designers have codified design systems that the agents follow. When a product matures, it graduates to a dedicated product manager, engineering team and designer, and the structure starts to look like a conventional product team. Levie observed that corporate innovation labs often vanish for six to twelve months at a time. Labs works, in his view, because its teams check in far more often and can show working software at each checkpoint.
Levie also said he had started to see models propose better product ideas than people would, because a model can borrow from dozens of domains at once. Krieger said he sees that at the feature level, and inside Anthropic as a habit of asking what another team did with a similar problem.
Find the people who get it, then keep a list
Asked for advice to technology leaders, Krieger offered two points. First, every organization, Anthropic included, has people who get far more from the models than their peers. He said the hard part is finding them and putting the most powerful tools in their hands, because the value they produce is disproportionate. He separated that group from the usual top-performer list. The two overlap without matching.
Second, he urged leaders to build their own sense of what the models can do. He recommended keeping a personal list of what AI still does badly, so you notice the day an item comes off it. He keeps such a list himself. A short census of who already works this way finds the first group, and a dated list of what the models still get wrong gives your AI policy a trigger for review each time an item drops off.
Since the November 19 listing, public investors hold a stake in whether this way of working spreads beyond Anthropic. Krieger's account shows the end state the company is running inside its own walls, with engineering in the chat window and people spending their time with customers.
If your engineers could prototype in an afternoon what used to take six months, would your organization have the customer access and the approval paths to use that time, or would the queue move from the developers to the approvers?
Find the people who already work this way before you buy more seats.
Ortutay, Barbara. "Anthropic Prices Its IPO at $83 per Share, Valuing the AI Giant at $1.5 Trillion." PBS NewsHour, Associated Press, 18 Nov. 2026, pbs.org/newshour.
"Anthropic Prices IPO at $83 a Share, Above Projected Range." TechCrunch, 18 Nov. 2026, techcrunch.com.
"Anthropic Shares Close 17% Above Their IPO Price on First Day of Trading." The Next Web, 20 Nov.
