Why Claude Still Needs a PM: Anthropic's Product Leaders on Judgment, Agents and Parallel Bets

Moderator Dan Shipper opens by noting that a year ago AI was said to make PMs obsolete, and now the vibe is that everyone needs to become one. Ami Vora answers that the job of product has always been a bridge between real human problems and the technology that can solve them. What has changed is that the technology used to shift every five or ten years and now seems to shift every two months.

Her approach is to stay obsessed with the human problem, which changes slowly, while staying fresh-eyed about the technology and throwing away much of what used to work. She likens the moment to rewinding the clock to when product had no definitions: you act as a general problem solver, keep going until you hit a wall, then use scaffolding to get through it.

Mike Krieger tells a story from his move to an individual-contributor role earlier in the year. When a project was close to shipping, a PM lead told him it really needed a PM, and he wondered whether Claude already had it covered. After the PM joined, the glue work around customer success, safeguards and keeping users in mind got done, and he messaged her a week later to say she was right.

Krieger says that because teams can move faster, the PM has to be operationally excellent beyond what was needed before, so builders can stay heads-down. Asked why a more advanced Claude cannot do the work, the panel says Claude does catch cross-organization issues through internal search, but is not yet a convenor. Vora describes the human archetype of the convenor, who brings the Claudes and people together, and jokes that the real moment will be when Claude schedules a meeting unprompted.

Vora adds that tools make it possible to build many things quickly, which expands what could be built while the information to decide is thin. That makes judgment about what to build and relentlessness in making it happen more important, because there are so many forking paths. Krieger compares each model release to looking beyond the horizon: there is neither utopia nor horror, just a new landscape with more to navigate.

Vora says she spent years learning to put herself in a user's mindset and decide, for example, where a button should go, because building was expensive. Now, she says, it is faster to build three versions and try them, so a skill she prided herself on matters less. She ties that to the innovator's dilemma at a personal level, and says she has never worked at a model company and has to try new things every couple of months.

To cope, the speakers emphasize raising your tolerance for change and talking openly about how hard it is, since everyone has to adapt. Vora says to frame the chaos so it feels safe rather than mapping out exactly what will happen, which locks you out of trying new things. Krieger adds that chaos needs clear leadership: in labs, a bet lead is the directly responsible person who decides whether to double down or wind a bet down.

Shipper asks how software changes when both humans and agents use it. Krieger sketches a timeline from an AI sidebar disconnected from the product, to AI-powered features, to agent-native design. His summary of the idea is that everything a human can do in the product an agent should also be able to do, and he says few products, including some of Anthropic's own, fully get this right yet.

When it works, he says, emergent behavior appears: the agent pieces together actions that were never connected before and can offer new ways of doing things proactively. The next stage, in his view, is interfaces that the agent itself can change. For a complex project with at least four independent work streams, his team had Claude monitor progress and build the UI that the TPM and then the whole team used. If they did not like the display, they could change it together with Claude.

He says that much of the software Anthropic uses is now built and maintained by Claude, which raises questions about who can update data and how to track its provenance. Building blocks, predictability and design systems are still wanted. For a company with an established SaaS app, he advises building the right primitives, a shared layer that agents and a REST API both go through, and evolving gradually, for example with a malleable personal landing page, rather than blowing up the whole UI. You can tell, he says, when an agent feels bolted on.

Asked how Anthropic serves everyone from casual users to large enterprises while pushing the frontier, Vora says the approach is to meet people where they are. It is too early to know what works, so the team should try many things and later fold what works into something stable. Her own background favored simplicity and predictability, but she now encourages the team to take many shots on goal and would rather have five overlapping products than one over-constrained one that never had a chance.

To keep parallel bets from feeling demoralizing, she says to name that openly: some teams will not be set up for success, nobody knows the answer, and everyone is on the same team. Krieger adds that a manufactured B team does not work, and that excited teams should explore. He says the main enabler is shared infrastructure: chat and co-work once had different memory systems, MCP and file storage, so a foundations team works on getting memory everywhere.

On signs that something is working, Vora says it is the usual product-market-fit questions: do people use it, like it, come back and talk about it. Krieger explains that for projects that fail only because models are not ready, labs will park them. The first computer-use product inside Anthropic in 2024 was so poor it helped neither automation nor education, but it was re-run in an eval harness with each model, and the team recognized a leap when it suddenly succeeded.

Shipper calls the risk capability blindness: trying something once, never trying it again, and deciding models will never do it. He polls the audience on their use of the latest models and says he often hears complaints from people who have only tried an older version. Vora says this is why it is important to be willing to throw away everything, or you assume the last thing that happened will stay true forever.

Shipper then asks how to turn successful parallel experiments into a coherent product without a pile of tabs or separate apps, admitting nobody has the answer. Vora agrees it is not fully figured out and points back to primitives, so the user does not carry the cognitive load of choosing tools. Krieger recalls the power law from Instagram, where the main feed was about 80% of use, and says half the job is saying no to another sidebar item while proving ideas out before folding them in.

Looking a year ahead, Vora hopes people will feel they can build, get answers and start businesses much more easily, with more small teams and solo empowerment. Krieger wants to close the gap between what models can do and what most people use them for, so that it is not just the most advanced software engineer with a multi-agent setup but also researchers and small-business owners. The art, he says, is letting them check in at the right places and do long-horizon work without feeling disempowered.