Why Claude can’t be your PM (yet) | Anthropic CPO Panel
In a Nutshell
The core message is that PMs are more essential than ever as AI accelerates development, because someone must still make judgment calls under uncertainty, coordinate people and systems, and keep user problems at the center. AI can't convene stakeholders, set meetings, or exert organizational influence, so the PM's role shifts from detailed user research to high-stakes decision-making and building "agent-native" products where agents can do anything a human can. Success requires relentless adaptability, tolerance for constant change, and willingness to discard old skills and habits as models improve every few months.
These notes were generated by AI and may contain inaccuracies.
A year ago, predictions suggested AI might make product managers obsolete, but now it appears everyone needs to be a PM. The core job of a product manager has always been to create a bridge between real human problems and the technology to solve them. While technology now changes every two months rather than every five to ten years, human problems remain constant.
Previously, the PM role lacked clear definition, so practitioners worked as problem solvers, hitting walls and either asking for help or figuring things out independently. The current approach involves returning to blurred roles and working iteratively, supported by structures when obstacles arise.
Mike shared an experience where he moved into an IC role and was building toward a launch. A PM lead intervened, stating they needed a dedicated PM on the project. After the PM joined, all communication, coordination, and critical tasks were handled properly. Mike later texted the PM: "Kat, you were 100 percent right." The PM ensured customer success teams could discuss changes, security measures were checked, and everyone worked toward correct goals.
Even with advanced tools, no one has unlimited power. A PM ensures everything connects properly and keeps user needs central. This role becomes more essential when teams work faster, requiring greater skill rather than diminishing importance. When hiring for teams, people who can perform these responsibilities are increasingly valuable.
The question of what a PM does that AI cannot do reveals that AI lacks organizational influence and the ability to schedule meetings. The concept of a 'convener' or coordinator role involves taking initiative to bring AI and people together to complete tasks. AI cannot currently set up meetings between people.
With advanced tools enabling rapid creation, teams must make right decisions based on limited information amid fast change. This requires determination to move forward in the correct direction among many possible paths. The need increases for someone with a plan, user access, and knowledge of people's problems to guide decisions.
Sound judgment and relentless effort become critical in the face of uncertainty. Each new model release creates a sensation of looking out to sea and wondering what lies over the horizon, only to discover more new territory ahead.
Product leaders previously spent extensive time understanding user mindsets and how products fit into their lives because building and bringing products to market was very expensive. This included detailed discussions about button placement, interaction patterns, and phone usage contexts. These skills are no longer needed because creating three versions and testing them is now much faster and easier.
The innovator's dilemma applies personally when someone becomes so skilled at one task that they continue doing it repeatedly, even when another approach becomes more important. Product managers must shake up what they knew about themselves and try new approaches, even when uncertain about mastery.
Adaptability and tolerance for change are essential. Judgment and relentless effort remain critical. Teams benefit from open discussion of these challenges to avoid feeling isolated in constant adaptation. The goal is bringing chaos into structure so it doesn't feel exhausting.
In uncertain times, people tend to predict exactly what will happen as a control mechanism, but this prevents trying new things. The approach involves presenting chaos in a way that feels safe and workable, acknowledging that emotions are involved and accepting that adaptation is difficult.
Clear leadership and accountability remain important alongside chaos. Labs designate "leads" for initiatives called "bets." The Bet Lead decides whether to push forward or wrap up projects, considering team size adjustments. This role gains importance as pace increases and situations change rapidly.
The industry is moving from a world where only humans used software to one where both humans and agents use software, sometimes through delegation and sometimes through collaboration. The evolution timeline shows AI initially isolated in sidebars, then driving entire features, moving toward agent-native architectures.
The principle states that anything a human can do, an agent should be able to do. Few products have achieved this fully, but successful implementation enables new behaviors where agents collaborate on previously impossible tasks. The next step involves interfaces that agents can change themselves.
Anthropic implemented a complex task with four independent work streams by having AI monitor everything and create a UI allowing the TPM and team to understand project status. The system allows changes through AI collaboration rather than being limited to internal team or third-party creation. Most software at Anthropic is now built, maintained, and developed by AI.
Questions arise about who can update information, whether changes come from human clicks or AI working behind the scenes, and how to confirm information source or provenance. The trend points toward making nearly all software function this way while maintaining building blocks, predictability, and design systems.
For companies with scaled SaaS apps considering adding agents, the recommendation is building the right primitives into the product as an infrastructure layer. Products should route everything through a common pipeline usable by both agents and REST APIs. Companies can start with side panels before moving to fully generative UIs or personal landing pages.
Anthropic balances pushing boundaries with providing clear, consistent experiences for users who prefer stable interfaces. The approach involves connecting with everyone based on their location and recognizing the early stage of understanding what works. Continuous updates are necessary because both models and usage patterns change.
When exploring new products, extensive testing is necessary because true boundaries haven't been discovered. Products that work should be made understandable for those wanting stability. The approach favors having five similar products that work differently over one overly restrictive product that never reaches the market.
To avoid confusion from parallel experiments, teams must be passionate about their product directions because unenthusiastic teams produce poor results. Different teams can pursue different great ideas simultaneously. Infrastructure improvements now allow memory and data sharing across products rather than isolated systems.
A foundation team addresses complex infrastructure issues so innovation teams can build without creating isolated experiences. Open discussion is essential, acknowledging uncertainty and framing the process as a shared journey where everyone learns together.
Product-market fit indicators include usage, satisfaction, return visits, word-of-mouth, and actual problem-solving. The challenge involves determining whether something isn't working or if models aren't ready yet. Projects can be put on hold and retested with new models rather than cancelled prematurely.
The computer-based product was initially poor because models weren't advanced enough. Early testing showed it wasn't good for automation or education. The project was put on hold and retested with each new model using an eval harness. Version 3.7 showed significant improvement in computer usage, demonstrating success over failure in transcripts.
Blinding oneself to potential by trying something once and concluding it will never work with current models. Three months later, a new model can completely change everything and improve capabilities in previously unimaginable ways. Many audience members have used Fable, while fewer have used Astra, indicating the audience is at the forefront of technology. When people say certain capabilities are impossible, asking whether they have used the latest model makes a significant difference. The importance of being prepared to forget everything previously known, as whatever happened last will not remain true forever.
When conducting many parallel experiments, incorporating successful ones into the main product presents challenges. Options include creating separate tabs where functionality differs, or keeping experiments as separate apps that never integrate. No definitive solution has been figured out for converting parallel experiments into a consistent product.
Building a fundamental framework that makes products feel intuitive when used, based on what users expect without requiring excessive thinking. Current interfaces put pressure on users by offering tools and asking them to choose. Creating a basic framework so the system understands the user and the UI always feels intuitive. These approaches are cyclical based on how people react, and betting on lessons that prove real.
Strong power law patterns observed in tab usage, with the original feed accounting for 80% of usage. Even improving Explore might increase it to 10-15%, but the main feed remains where users spend most time. The journey involves identifying the main tab experience and making it great while serving as a gateway to other experiences. Half the job involves saying no to sidebars, additional tabs, or extra features and considering whether they can be included in the main experience.
Things are changing so fast that this represents the entire industry of product development. Next year, hoping people feel able to do more than ever before, create anything they want, get answers whenever and wherever needed, and start businesses. Greater empowerment of small groups and individual initiatives, with benefits extending to others through creators who influence people who may not directly use the tools.
The difference between model capabilities and what most people actually use them for represents a failure in product creation. The goal is to narrow this gap, which becomes more democratic by enabling not just skilled software engineers to access fully functional multi-agent systems, but also researchers working in parallel and business operators. The approach involves understanding different parts, monitoring the right places, and working long-term without discouragement. This represents real art, and success would bring the industry closer to this vision.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
