TL;DR
The short version
The org chart is turning into a set of loops — agents that take an input and drive it to impact for each function. The job of a leader is to build those loops, find where they get stuck, and save your own judgment for the leaps a loop can't make.
That's the frame from a16z's Anish Acharya on Lenny's Podcast: for a decade the question inside companies was "who owns this?" — the new unit isn't the person, it's the loop.
Built on Lenny's Podcast with Anish Acharya (a16z), hosted by Lenny Rachitsky.
Every function becomes an input-to-impact loop
An agent is a model running in a loop with tools, memory, and skill files. Coding already works this way: a bug report comes in, a repro gets generated, a fix gets written and reviewed, and if it's low-risk it ships and emails the customer — in about five minutes. Acharya's claim is that this shape spreads to every function: a loop per person, per job function, per business unit, up to loops that run large parts of the company.
Take a growth team. The old version was a room, a brainstorm, a spreadsheet of experiments, and weeks of prioritization. The loop version generates every variant, measures every variant, ships the one that hits statistical significance, opens a long-term holdout, and starts the next test. Sales, support, legal, and product marketing follow the same pattern: the administrative work around the core job gets handled, and the person is left with the part they're best at.
The operating move is to stop asking "who does this task?" and start asking "what's the loop, and where does it break?" Acharya's sharpest example is Kavak, which sells used cars online: an agent handles each customer, and when it gets stuck it calls a human. The human coaches it through — and the agent captures the trace and learns, so next time it doesn't need the call. Every block is a knowledge gap or a data gap you close once.
The loop climbs the hill; you find the next one
A loop is a hill-climber. It reaches a local maximum, then it plateaus.
You need somebody to actually help you land at the base of the next hill.
Anish Acharya, Lenny's Podcast
This is the boundary, and it's where the leader still lives: strategy, exceptions, and the genuinely new idea. Ask an agent to "make me a million dollars, make no mistakes" and it stalls, because it needs to be pointed in a direction first. Models are still weak at out-of-distribution thinking, so the human job compresses to picking the next hill and being right about it.
There's a quieter benefit. When every idea gets built and tested instead of pitched, the idea that wins isn't the one whose owner sells best to an executive. One executive told Acharya his team hadn't laid anyone off; instead "2 years of roadmap happens in 3 months," and the hard problem became knowing what to add.
Match the model to the upside
Not every loop needs the best model. Acharya splits work by upside. For unbounded-upside jobs — drug discovery, research, sales, engineering — pay for frontier intelligence, because one more increment of quality can be a trillion-dollar outcome. For bounded, verifiable work — closing the books, routine legal — use cheaper open-weight models, because you can't close the books 100x better. Past a certain threshold, extra intelligence on a bounded task is simply waste.
Stop optimizing only for productivity
The contrarian point is the one worth sitting with. Acharya thinks the industry has the goal wrong: people don't only want to save time — more of them want to spend it. His line is that we spent 40 years building technology for better spreadsheets and nothing to extend our soul.
For product leaders that reframes the roadmap. The loops that grow a business — close the ticket, ship the variant — are the obvious ones. The under-built loops are the ones aimed at a person's life: improve my health, make me a better friend. He frames this as a product-design problem, not a model problem.
The minimum viable practice underneath all of it is small: ship something every week, even something unimportant, so you build the intuition to see where a loop should go. Build the loop, watch where it breaks, and keep your own attention for the leap it can't make.
