Be a Willow Tree: Julie Zhuo on Managing Agents, Teams of Builders, and Data That Diagnoses

Host Lenny Rachitsky opens by suggesting that almost everyone will soon be a manager, because agents are being woven into daily workflows. Julie Zhuo, author of *The Making of a Manager* and founder of the data startup Sundial, agrees completely. In her view management means having an outcome, a north star, and figuring out how to use your resources to reach it.

She breaks the job into three levers: people, purpose and process. With agents, "people" becomes models, which have different strengths and something like different personalities, so you have to develop an intuition for which one fits which task. You also have to decide which tools an agent can reach and what context and high-level instruction it gets.

The skill she would double down on first is defining success. Alignment is hard in big companies because people carry different pictures of what a good outcome looks like, and "build an amazing product" is far too vague. Boiling a goal down so that an agent can tell success from failure is, she says, a big part of the game, and it is exactly why evals and metrics exist.

Rachitsky raises the flattening of organizations, with ICs again and middle managers being cut. Zhuo sees the real promise of AI as individual empowerment. Where someone once needed to hire a designer, a coder and an analyst, AI can now take a beginner to roughly the 60th or 70th percentile in many of those skills, which is enough to unlock a lot of work.

From that she concludes that teams can shrink to two people from any traditional discipline, and that we can "call ourselves builders". At Sundial the company has eliminated roles it once expected to need, including product managers. Her reasoning is that when a PM is present, an engineer facing a product question defaults to handing it over, whereas without one the engineer owns the user experience.

She is careful to say that not everyone must be good at everything. A team still needs someone with product and user sense, whether that is a designer, a data analyst or an engineer who happens to be strong there. The same applies to front end and back end: she asks engineers to use AI to stretch into the other side, with a specialist reviewing the code. Projects take longer at first, but the investment pays back in more well-rounded people.

Engineering is the most accelerated function at Sundial, but Zhuo also sees designers, engineers and a hybrid "product science" team prototyping and picking up each other's work. When she asks engineers to do analysis, they often say they don't know how. Her answer is that ChatGPT and similar tools are now better teachers than pulling a colleague aside.

She feeds a 12-week online curriculum into ChatGPT and asks it to customize a program around how she learns best, which for her means lots of analogies and explain-it-like-I'm-five examples. Different teammates learn differently, so a personalized tutor accelerates each of them. In her view this is more underused than the coding tools themselves.

One engineer told her that he understood the algorithms for root cause analysis but not when or why a company would need them. That broad contextual question is ideal for a general-purpose model. She also uses the tool to test herself by retelling a concept in her own words and letting it critique her, which it does politely before eventually saying she was completely wrong.

Asked how AI companies use data, Zhuo says she doesn't think many fast-growing ones use it well yet. Historically a company reached a hundred million users only after years, which left time to build logging, hire a growth and data team, and set up observability for the business. Today companies with a handful of people and hundreds of millions in ARR skip all that and get by on good instincts.

The trouble, she says, is that growth does not last forever, and when it stops everyone asks what happened. A business that has been instrumented well can find root causes, spot slowdowns earlier and even predict them. A business without that visibility ends up scrambling, and that is usually when companies start investing heavily in data.

She adds that each technological shift demands new analysis methods. Mobile made sessions per day and session length important, whereas conversational products cannot be measured by tab clicks. Teams need an LLM or machine learning model to bucket user intent, and they have to work out whether a one-question chat means value or failure. Rachitsky adds that ChatGPT's team found early popular use cases partly by watching viral TikTok comments.

Rachitsky points out that designers often resist experiments and data, and asks what they miss. Zhuo says she and her co-founder repeat one phrase: diagnose with data and treat with design. Data will not tell you what to build, but it can show whether you have a retention problem and where the opportunity sits, and then a creative process has to find the solution.

She defines data broadly. It includes A/B tests, but also what went viral on TikTok, what people say on X, and customer interviews, which AI now makes easier to synthesize. All of it is a story that reflects reality better than the one we want to believe, such as "people love us".

She also sympathizes with designers: choosing which numbers to look at and deciding whether a 5% rise is good are acts of interpretation, and you can't A/B test your way into a great product. But she says it is not data versus design. Every great designer she has met is obsessed with learning what users really think, and Rachitsky agrees from his own experience.

Asked about the biggest change in management, Zhuo says it has always been a manager's job to manage change, but the rate of change is accelerating. Nobody knows where AI will be in two years or whether AGI is five years away. Many people also fear that tools are getting good at what they do, so managers must handle that uncertainty on top of learning to manage AI.

That makes communication, feedback and compassion more important. Leaders have to tell teams that the product and the way of working must both change, without freaking everyone out. Her metaphor is the willow tree, which survives storms because its branches are flexible, so management today means being "sturdy while being flexible".

Rachitsky recalls Marc Benioff's answer to the same question, which was to say "good" and treat change as opportunity. Zhuo agrees that leaning into excitement beats living in fear, but warns against pretending that nobody is upset. Admitting that change is hard and that chaos will come, while keeping the big picture in view, is part of leading through it.

For timeless advice, Zhuo points to managing yourself, the chapter she wanted to put first in her book. Her key idea is dimensionality: each person has infinite dimensions, from ax throwing to podcast hosting, and each is a single facet. If you accept that, feedback on one dimension doesn't threaten your identity or worth, and conversations become more objective.

She also believes every strength is a weakness. Early managers praised her for being thoughtful and frameworks-driven, yet criticized her for being quiet and slow on her feet in live discussion, which are two sides of one trait. Mastery, she says, means working on the opposite skill, for example by saying "I don't know how I feel yet, but here's my current thinking", without rejecting what you are good at.

How far to push on weaknesses depends on your goal. An individual contributor who wants to deepen a craft may not need to manage, as long as a path exists to the income and impact they want. Suffering comes when goals are out of sync, like wanting a mansion while spending your time perfecting an omelet. She also advises asking your manager what success looks like for the team and what a promotion would take, which turns the relationship into a collaboration rather than an adversarial one.

Zhuo is passionate about feedback and thinks teams undervalue it by waiting for six-month review cycles. What matters is how quickly a team improves, and a team that gets 1% better every week will beat one that improves 1% a month, even from a lower baseline. She treats feedback like data: a hypothesis about yourself tested against reality.

People are often out of sync with reality in both directions, she notes. Overconfidence and imposter syndrome are both biases, and we need others to reflect back what we can't see. Her image is someone telling you that you have a leaf on the back of your head, a gift you can't give yourself.

On delivering hard feedback, she gives three tactics. First, establish early that you both want to help each other grow and that all feedback is open, which she says covers about 60% of the difficulty. Second, check your intention, because wanting to be right or to punish someone will show. Third, say out loud that you are nervous and that you value the relationship, since vulnerability humanizes the conversation.

Zhuo says managers often hold an adversarial story, in which better outcomes mean pushing people to do things they don't want. She prefers a win-win mindset. The first time she had to tell someone they didn't belong on the team felt horrible until she saw that someone in the wrong role is already unhappy, and that helping them find a place where they can grow is respectful.

Rachitsky notes that this only works if you believe it, and Zhuo agrees. Another trap is lack of conviction. Middle managers often feel like soldiers executing a vision they didn't create, but if you don't believe in a strategy it is hard to convey purpose, and she can't recall a project turning out wildly successful when its manager doubted it.

Her advice for disagreeing with a CEO's priority is to engage in dialogue and decompose the proposal into hypotheses. You might believe some and doubt others, and then you can propose a small test, such as trying a lemonade stand only in one market before opening them across fifty states. That lets you tell your team honestly that you are unsure but committed to finding out.

In the show's AI corner, Zhuo describes building gifts for her children. Borrowing an idea from Eric Antonow, who put a microphone and speaker into a parrot connected to ChatGPT voice mode, she made a talking raccoon for her six-year-old. For her middle son she built a Replit app that writes personalized video game parody songs from a Spotify link and some inside jokes, which she will sing herself.

Her contrarian belief is that "there's infinity in every direction", so if something seems boring, the problem may be that we haven't yet seen its richness. Staring at a wall can be miserable for her kid and wonderful for a monk. She finds this more empowering than feeling like a victim of circumstances, because it implies a skill you can grow.

In the lightning round she recommends *Zen and the Art of Motorcycle Maintenance*, *Conscious Business* and *Good Inside*, and praises the Matic robot vacuum and the Limitless pendant, in which she is a small investor. Asked what she teaches her kids for an AI world, she says emotional regulation. She fears AI will make it too easy to shortcut hard feelings and relationships, and wants them to pick challenges deliberately rather than treating AI as a shortcut.