The guest's optimism starts with limits on an intelligence-only story
Source 03:09In this Lenny's Podcast conversation, a16z investor and former product builder Anish Acharya explains why he expects companies to become collections of agent loops. His picture still includes people choosing direction, handling exceptions and building products that consumers want to spend time with.
He pushes back on the idea that missing the latest AI tools inevitably creates a permanent underclass. He points to a varied technology ecosystem, the pace at which organizations adopt new tools and problems that are not limited by intelligence alone.
Those are his arguments about the future, not proof that disruption cannot happen. His example is a business such as a pizza chain: access to a data center of highly capable thinkers would not automatically remove every physical or organizational constraint or give one company the entire market.
Giving employees a tool differs from redesigning the company
Source 08:43Acharya objects to assuming that everyone else's work is easy to automate while treating one's own work as special. He says employees often want more leverage and can learn the tools when given the opportunity.
He describes Kavak's internal training program, including mechanics learning to use AI and participants shipping a production agent after a six-week course. The example supports his interest in teaching people, rather than viewing them only as tasks to replace.
He then distinguishes introducing AI into existing job functions from reorganizing a company around it. His historical analogy is factories adopting electricity: changing the power source and redesigning the whole production layout are different levels of change.
A business loop connects action to a result
Source 11:50Coding supplies a concrete starting point. A bug report arrives, a reproduction is generated, a fix is made and reviewed, and release depends on risk. The loop can continue to a customer notification that the reported problem has been addressed.
Acharya asks what the equivalents are across marketing, sales, support and legal work. He imagines loops at the level of a person, a function, a business unit and larger parts of a company, with the outputs of smaller loops informing broader decisions.
He does not describe most organizational work as ready for full autonomy. Sales, support, strategy and exceptions remain important human activities in his account. He is especially skeptical that progress in mathematical reasoning automatically demonstrates the ability to invent a good business direction.
Optimization reaches a local maximum
Source 14:42A growth team traditionally gathers possible experiments, prioritizes them, builds, ships and measures them. In Acharya's loop version, more variants can be generated and measured, successful changes shipped and longer-term effects checked with a holdout.
The limitation is the hill the process is climbing. A loop can improve within a direction until it plateaus. Human intuition is then needed to suggest a different direction, or what he calls the base of the next hill.
At 15:30, the source slide illustrates humans choosing a new hill and agents testing, shipping and measuring within it. It is a conceptual model, not measured performance data.
The discussion of asking an agent to make a million dollars makes this point explicit: it still needs a useful direction. More repetitions of an underspecified objective do not by themselves supply the missing business judgment.
Treat a stuck agent as a question about missing context
Source 19:02Acharya offers a practical diagnostic: when the model makes a choice you would not make, ask what you know that it does not. The gap may be information or context rather than a lack of effort.
His Kavak example is an agent that calls a human when it gets stuck. The person coaches it through the situation, and the resulting trace can help improve future behavior. The intervention both unblocks the immediate task and creates an example to learn from.
That is the feedback connection in his organizational picture. Giving an agent a task is only the beginning; people still have to notice where it fails and provide the knowledge, data or direction needed to improve the loop.
Model cost depends on the upside of a better answer
Source 21:55The conversation also asks when to use frontier models and when narrower, cheaper systems make sense. Acharya frames the trade-off around price, performance and how much value another improvement could create.
For a problem with unusually large upside, paying more for stronger reasoning might be worthwhile. For a bounded task that must simply be completed correctly, paying indefinitely for marginal intelligence may not make sense. He discusses using different model families and specialized configurations within the same organization.
He also complicates his own distinction. A support call that looks routine might contain the first clue to a company-changing problem. Verifiability alone does not settle the choice either: both a bounded accounting task and a high-upside discovery can have verifiable results.
The account therefore contains competing cases for broad intelligence and efficient specialization. It is not a fixed ranking of departments or a universal purchasing formula.
Build something to learn how models differ
Source 26:38Acharya says he tries to ship something with each new model. Actual projects reveal differences that are hard to see when models are treated as interchangeable labels on a leaderboard.
Keeping one or two projects alive gives him a recurring place to try new capabilities, instead of inventing a fresh task every time. The project can be playful or unimportant. Its role is to make experimentation concrete.
One family experiment also exposes the weakness of a metric. After he tied an automated screen-time system to positive behavior, his son replayed a recording of himself saying “I love you” beside the microphone. The story is an example of a measure being gamed, rather than a recommendation to use that system for parenting.
Consumer value includes connection, progress and fun
Source 32:18Acharya argues that many consumers want to spend time, not only save it. Entertainment and social products address needs such as connection, enjoyment and a sense of progress. He wants more AI product work in those directions.
He identifies several barriers: model costs, interfaces that require too much initiative and a focus on productivity rather than these other needs. His description of a space between chat and a feed is a product-design question about helping people find value without making them specify everything from scratch.
He groups opportunities around coding agents as general problem-solving tools, simpler personal agents, and creative or entertainment products. His broader idea of ambition also includes family, local services and making music, rather than only starting a larger company.
Durability can emerge from product use
Source 54:54Asked about moats, Acharya says they are often discovered rather than fully designed in advance. He describes products that begin with strong craft and engagement, then develop advantages through use and accumulated experience.
He also notes that familiar sources of durability were never merely about software being hard to write: network effects, scale, brand and proprietary resources still matter in his account. A large product idea is often supported by many small choices a competitor cannot see just by copying the headline concept.
His preference is to listen to customers and observe momentum rather than demand only an elaborate theoretical moat story. That is his investing perspective, alongside the continued relevance of those older business mechanisms.
Distribution remains difficult, even with a more ambitious product
Source 59:26Lenny raises the problem of reaching people when many products launch every day. Acharya emphasizes word of mouth and building channels around genuine interest, since established networks are not necessarily welcoming places to build another network on top.
He also challenges founders to imagine a much more valuable version of the product. Thinking about what could justify a very high price is, for him, a design exercise that can expose limits in the team's imagination.
The exchange keeps both concerns visible: the interviewer asks how a product first gets attention, while the guest argues for stronger craft and ambition. A better product and an actual distribution route are related questions, not interchangeable measures.
A weekly project can be small and still teach something
Source 69:35Acharya's closing advice is to make more things and use a project as a place to learn the models. His suggested cadence is shipping something each week, without requiring it to become a company or remain useful forever.
He gives the example of making a Mother's Day presentation from old messages and photos. It was a personal gift, not a startup, and still offered a concrete experience of creating with the tools.
The point is to build intuition through work that reaches some kind of completion. A small finished project can teach a capability, reveal a limitation or bring someone joy even when it has no broader audience.
Adapted from Lenny's Podcast, “Why companies are becoming a series of loops,” with Anish Acharya. Predictions, business anecdotes and product judgments are attributed to the conversation. Sponsor and lightning-round material outside the main line are omitted.