从制品到有机体：微软Asha Sharma谈会学习的产品和智能体社会

Sharma says that after OpenAI and the large foundation models arrived, there was an explosion of proprietary, open and frontier models that were more efficient and increasingly domain-specific. Models can now call tools and functions and take action, which is giving rise to a new kind of product. Instead of shipping a static artifact and watching a dashboard, a team's key measure becomes its metabolism: how well it ingests data, digests a reward model and creates an outcome.

Because the models are so capable, teams can tune them toward outcomes such as price, performance or quality, and the products improve with more interactions. She calls these living organisms and says this loop is the new intellectual property of every company. When Lenny Rachitsky links this to Cursor's data moat, she says using your own data is the best route, but data can also be synthetic, provided you design rewards, roll them out and A/B test rigorously.

She has never seen this run as a single loop; it is several parallel tracks, like assembly lines. She expects most software to become model-forward, and Rachitsky adds that ChatGPT's interface could one day reshape itself for each user. Sharma agrees that your experience should look different from anyone else's, and different again for agents, power users and newcomers.

Sharma describes a pattern across the companies doing well with AI. First everyone becomes AI fluent by using a copilot day to day, so no one is afraid of it. Second, they take an existing process, like customer support or cutting fraud resolution from 15 days to 10, and apply AI to it. Third, once the impact shows up in the P&L, they use AI to drive growth, for example through retention or entirely new categories.

Companies fail when they do AI for AI's sake, launching many projects at once with no blueprint, no real investment and no measurement, observability or evals. She also notes that around 70,000 enterprise AI tools launched last year, so enterprises should bet on a platform layer that lets them swap pieces in and out. Her phrase for it is to build for the slope instead of the snapshot.

Builders are changing too. Sharma says each technology shift created new roles, but now she sees the polymath and a renaissance of the full-stack builder. A typical organization has many functions and layers, with about 500 touch points to launch a product, which is too slow when hundreds of new models and technologies arrive each week. She says even 50-year-old enterprises are starting to work with full-stack builders, and Rachitsky notes the Venn diagrams of PM, design and engineering keep converging.

Her slogan is "the loop, not the lane." Whatever your function, you need to understand the cost of the product, the reward and system design, and how the interface works for people and agents. As she puts it, feedback becomes continuous, observability becomes culture, and functions blur. One example is Microsoft's Dragon product for physicians, where synthetic fine-tuning gave way to 600,000 expert-annotated patient-physician interactions and acceptance reached about 83%, driven by a small cross-functional group.

Sharma says everyone's instinct is a GUI, but history points elsewhere: databases moved from the desktop to SQL, and cloud moved from consoles to Terraform. A stream of text connects well with LLMs, so she argues product makers should think more about how something composes and how an agent will read it than about the canvas. That also helps with scale and collaboration.

Asked whether the future is terminals like Claude Code or agents, she says nobody knows, but terminals feel great because they fit the text stream an LLM works with. She expects humans to keep committing code through IDEs, GitHub Copilot or new environments, while agents work with each other. On chat, she says conversation is a powerful interface but people also live in email, docs, Word and PowerPoint, so chat will be important but not sufficient.

Rachitsky relays Nick Turley's view that ChatGPT is in its MS-DOS era, with a Windows-like version to come. Sharma finds that smart and says every company should bring AI to where its users already are, and ChatGPT's users are in chat. The open question, she says, is whether it will always be chat.

Sharma says we are only starting to scratch the surface of an agentic society. She expects the marginal cost of a good output to approach zero, which will create exponential demand for productivity, and agents, both embedded and eventually embodied, are how work scales. You can already assign a pull request to Copilot or build an agentic software development rep for lead generation, and she thinks there will be far more agents than today's software.

In that world tasks and throughput matter more, layers shrink, and the org chart starts to become the work chart. Rachitsky imagines a conversion-onboarding team made entirely of agents prompted to drive conversion. Sharma says people will stop thinking in terms of hierarchy and reporting up and think more about task-based opportunities, with humans still deciding how AI is used. The key operational questions become how new tasks get routed, how you observe agents and how you fine-tune them when they get things wrong.

On reviewing a thousand agents' output, she says the same loop applies: fine-tuning, self-healing observability and good evals. She also points out that device, policy and group-access management for billions of people are solved problems, so not everything has to be reinvented. Her optimistic view for employees is that each can bring their own stack of agents to work, much like bringing your own device, and become more skilled.

In her own work she uses agents for summarizing live-site incident bridges, where 15 people may be talking, and for prototyping, where chatting in plain language can yield something more expressive than code. She also muses about whether strategy docs will survive. Her view is that consequential products need human judgment, vision and iteration, and that docs will fade into applications and other artifacts in the productivity suite.

Sharma says everyone is still figuring out planning, and it is harder in a larger organization. Microsoft's product teams used to plan in six-month semesters, but she says it is hard to predict that far out now. Instead they ask what season they are in, defined by secular changes in the industry or among customers: the early prototyping of AI, then models and reasoning models, and now the rise of agents. A season can last a year or only three months.

Within a season the team keeps a shared north star, loose quarterly OKRs and squads with four-to-six-week goals, and it leaves slack in the system for the unplanned. She says more than 15,000 customers have built agents on the Azure service, and corrected herself that the number of agents is in the millions. The priorities now are alignment, accountability, observability and evals, plus building blocks such as longer-running tool-calling loops, a breakthrough she credits to Manus, and memory.

What she wishes she had known on taking the role is that what makes a great platform is what makes a great product, and it often lives in invisible work. At Porch Group, the most important thing was a matching platform linking professionals, service types, zip codes and homeowners. WhatsApp won on the phone book, speed and reliability, and end-to-end privacy, not stickers or stories. For Microsoft's AI platform, she says it comes down to data residency, availability, reliability, the right tool selection and retrieval of knowledge.

Asked what she has learned from Satya Nadella, Sharma says optimism is a renewable resource: his ability to generate energy and renew everyone's commitment to the mission is a core part of the culture, alongside the growth mindset. She says people have to work on something deeply moving that they believe will make the world better. What drives her personally is AI's potential in healthcare and workforce productivity, including the hope of finding a way to treat the kind of cancer her mother has.

She looks ahead to a new wave around reinforcement learning, which she believes will be one of the most important product techniques of the next few seasons. She expects as much money to be spent on post-training as on pre-training, and eventually more, and says around 50% of developers are already fine-tuning according to surveys. New infrastructure, platforms and companies will form around this part of the stack.

Explaining the shift simply, she says building a foundation model needs enormous compute, scientific expertise and capital, while a rich set of good models already exists. Adapting an off-the-shelf model with reinforcement learning or fine-tuning on your own, bought or synthetic data gives more economic leverage and more control over taste. She favors a model-system approach with diversity over one model to rule them all, since models differ in latency, thinking time and strengths.