NLX Is the New UX: Microsoft's Aparna Chennapragada on Designing for Agents

Chennapragada does open mics and has performed a few shows, after a visit to an open mic with her son, who sings, led to an accidental discovery that she enjoyed it. She is working on a set about AI, tech and Silicon Valley. She jokes that she never expected to be serious about a funny business.

She argues that comedy and product building both have a version of product-market fit. An open mic is a live experiment with clear micro feedback, sometimes harsh, and she says that resilience is a valuable skill when the first version of an ambitious product is not quite there. She cites Reid Hoffman's advice that if you are not embarrassed by the first version, you launched too slowly.

One joke from her set compares AI chatbots to men: they hallucinate a lot, are not yet reliable, and sound confident even when they do not know the answer.

Coming from consumer companies such as Google and Robinhood, Chennapragada says the first difference at Microsoft is that every enterprise use case is really two: the feature must work well and it must be governed. Even sharing a link to a document should be frictionless and also secure and auditable. She says people often fall into the trap of ignoring one side or crippling the user experience for the other.

The second difference is what she compares to Van Damme doing the splits between two moving vehicles. The tech cycle is the most compressed she has seen, measured in weeks and months, with a large intelligence overhang, while productivity habits and change management move slowly. The future, she says, is unevenly distributed even within companies.

Her advice is not to hold back early adopters. Microsoft runs a Frontier program that gives experimental features, such as a deep research agent post-trained for work, to people who want to try them, while the broader rollout stays careful. She wants to institutionalize her personal habit of living a year in the future by asking what a team of three with lots of compute and AI tools would look like.

Chennapragada takes a practical view of agents: underneath they are tools, built on stochastic models rather than deterministic programs. She sees an evolution from apps to the assistance era, which she associates with Copilot, where the human drives, toward more delegation as reasoning improves.

Her example is a researcher agent she asked the night before to look at everyone in an upcoming leadership meeting, find their views on her roadmap and suggest how to pitch it. She says the value was not just time saved but new insights, which she calls superpowers.

She summarizes the three traits: autonomy as a spectrum of what you can delegate, complexity beyond one-shot requests such as a prototype of an augmented reality app, and asynchrony, where it works while you are away. She later reframes the third as natural interaction, which may mean talking to the agent in a meeting and not only chatting. In her view these are product principles for good agents.

Chennapragada says NLX is the new UX, one of her pet topics. Graphical interfaces had to be explicitly designed and are rigid, while conversation is elastic, but elastic does not mean undesigned. She notes that a podcast and a Microsoft meeting are both conversations, and conversations have structures and invisible UI elements.

She names new constructs. The prompt is a UI element like a dropdown or menu, and for agents an editable plan is another. Showing progress is a third, and it needs balance: too verbose feels like watching a script run, too terse leaves users unsure whether the agent is on track.

Follow-up suggestions are another design problem, for example what to propose after an image comes back as black-and-white clip art. Too many annoy people, and too few waste a chance to guide them. She thinks showing the reasoning is partly a point-in-time need, because today's models still feel like black boxes, and that the right level will become personal.

Chennapragada says that today, if you are not prototyping to see what you want to build, you are doing it wrong, and she calls prompt sets the new PRDs, or demos before memos. Her second point is that time to first demo gets much shorter but the time to full deployment gets longer, because a flood of ideas raises the bar for scale. Her third is that blurred roles still need a few taste makers, or the result is a Frankenstein product.

She disagrees strongly with the idea that coding is dead. Programming has always moved to higher abstractions, and she suggests there will be far more software operators, but understanding computer science remains a way of thinking. At Microsoft, she says, the culture is shifting toward live prototypes for new products and features, though not for deep changes inside Excel.

On product management, she says someone who is just a process person has a question to answer, but taste making and editing become more important when ideas and prototypes are abundant. She sees less gatekeeping, so product leaders must earn their editing role, while engineers, researchers and designers can bring forward good ideas with an AI expert in their pocket. Her own use is asking what Satya Nadella would think of a pitch.

Chennapragada says there is a coiled-spring intelligence overhang, and product teams must rewire themselves toward what Shopify's Tobi Lutke calls reflexive AI usage. She built a Chrome extension that asks, on every new tab, how you could use AI for what you are about to do. She wrote it with GitHub Copilot in about ten minutes and has no plans to release it.

She says the real difficulty is updating priors. Models could not do many things a year ago, such as spelling in images, deeper reasoning or data analysis, so impressions from a few months back are stale. In her words, the baby grew up to be a 15-year-old in a month.

She sees an arbitrage for builders who drop the scar tissue from earlier failures and keep demanding more from the AI. Asked about working closely with both Satya Nadella and Sundar Pichai, she describes Pichai as calm and measured with complex ecosystems and Nadella as having a huge appetite for learning and the ability to zoom between strategy and detail.

Her counterintuitive lesson is solve before scale. In the solve stage, wide lurches are normal: she recalls Google Lens moving from plant detection to foreign-language translation, and says it looks like chaos but avoids prematurely climbing a local hill. She is also wary of grown-up metrics too early, since click-through or retention with a thousand users is false precision.

She points to voice assistants as an example: Alexa, Siri and Google Assistant promised you could say anything, but they were really good at a few things like timers, music and trivia, and she says you have to nail those before promising everything.

For zero-to-one products she looks for at least two of three inflections. The first is a technology step change, such as deep learning for Lens, speech recognition for conversational search or mobile for Robinhood. The second is a consumer behavior shift, like people taking pictures of everything. The third is a business model shift, such as the second-price auction for search, and for AI seat, usage or outcome-based pricing.

Lenny Rachitsky asks why Microsoft, with Copilot, VS Code and strong AI talent, did not own the AI coding tool market that Cursor and others are growing in. Chennapragada says code generation is a tool that unlocks many products, and she describes GitHub as a system: a repository, context, autocomplete, chat and an agent board, which she says is one of the fastest-growing loops with strong positive feedback.

She says companies running on code need a Swiss Army toolkit and not only autocomplete, chat or a supervised agent. Rachitsky summarizes that all roads lead to GitHub, and she agrees that code generation products are not all doing the same job.

On Excel, she recounts asking an original Excel product person why it endures and being told it is a programming tool for non-coders. She also points to the World Excel championships as proof that some tools have a learning curve but reward users with depth, and credits decades of attention to signals from people who depend on it daily.

Her pivotal career moment came at Google, where she spent about a year trying to make personalization work in search, and it did not matter as much as expected, so the team was disbanded. She then worked on Google Now, which pushed useful information such as when to leave for a flight. It taught her that she loves seeing around the corner more than scaling, that being early is the same as being wrong, and what a small group of talented people can do.

Rachitsky says the story combines his Failure Corner and Hot Seat Corner segments. Chennapragada says the work became one of the foundations of Google Assistant and now works better with Gemini, and that sometimes the intelligence takes five or ten years to arrive while other times a product is one or two clicks from great. She calls Google Now a first agent before agents, where the interface overshot the intelligence.

Today she sees the opposite problem: strong intelligence with an interface like an AOL dial-up modem chatbot. The area she most wants to explore is a co-working space where humans and agents collaborate, deciding what to delegate, what to inspect and how information flows between people with agents mediating.