High Output Management 2.0

Andy Grove wrote High Output Management in 1983. It’s a Silicon Valley staple and one of the best books I’ve read on how to run a business. His central insight was deceptively simple: a manager's output is the output of their team. You don't judge a manager by what they personally produce. You judge them by what they're able to get done through others. Leverage. 

Grove defines high-leverage activities as the things a manager should prioritize above everything else. Specifically, training, creating processes, and unblocking. Do those things well and everything downstream gets better. Do them poorly and no amount of personal effort compensates.

He also introduced the concept of Task Relevant Maturity: how much oversight you give someone should match their reliability on that specific task. Low reliability requires oversight and narrow scope. High reliability earns autonomy. You don't manage a seasoned operator the same way you manage a first-week hire.

Grove wrote this book in a world that looks dramatically different than today. His focus was predominantly on people. Then I heard Andrej Karpathy describe AI agents, and it clicked: Grove may have accidentally written the management manual for the AI era four decades early.

Shortly after my last post, Sequoia dropped an interview with Andrej Karpathy about agent engineering. If you're not familiar with Karpathy, he co-founded OpenAI, got Tesla Autopilot working, and coined the term "vibe coding." He's about as credible as it gets in AI. 

And yet a few minutes into the interview, I could barely keep up with the points he was trying to make.

Stephanie Zhan from Sequoia had great questions but his responses were jargony answers like:

"The neural net becomes the host process and the CPUs become the co-processor."

But then he said something that stopped me:

"LLMs can easily automate what you can verify... when things are not kind of like in that space, you need to actually be in the loop a little bit and you need to treat them as tools."

That's Task Relevant Maturity. He just described it from the AI side without realizing it. Low verifiability — stay in the loop, narrow the scope. High verifiability — step back, let it run. Grove wrote that in 1983 about people. Karpathy just rediscovered it about agents.

Here's the irony: the people with the most to gain from AI agents right now aren't engineers. They're managers and operators. People who already know how to get things done through other people. They just haven't been invited into the conversation yet because the people holding the microphone keep speaking a language that signals, intentionally or not, this isn't for you.

Let me translate a few moments from that talk, because buried under the jargon is something worth hearing.

He said: "These are giant reinforcement learning environments... they end up basically creating these jagged entities that really peak in capability in verifiable domains."

He meant: AI is weirdly uneven. Brilliant at some things, bizarrely bad at others, and nobody fully knows which is which until you hit the wall. Just like with any new hire. It takes time to build confidence and delegate.

He said: "What is the piece of text to copy paste to your agent? That's the programming paradigm now."

He meant: The new way to get computers to do things is just describing what you want in plain English. That's it. Non-technical people have been doing this instinctively their whole careers. They just didn't know it now counted as programming.

He said: "Vibe coding is about raising the floor for everyone... agentic engineering is about preserving the quality bar. You're still responsible for your software just as before, but can you go faster?"

He meant: The tools changed. The accountability didn't. You can move faster now, but the standard you're held to is exactly the same. Faster execution with the same quality bar has always been what management looks like.

For the last few years, the benchmark in tech has been the 10x engineer, or one talented engineer able to do the work of 10 with the right tools. Claude Code and Codex have made that dream closer to reality. AI gave the best engineers an enormous force multiplier and the gap between great and average widened considerably.

With that framing, it’s understandable why engineering breakthroughs get all the attention. But Karpathy’s talk suggests we may be missing where the real leverage is shifting: management.

Agents don't just make engineers more productive. They make the Grove model of management scalable in a way it has never been before. A manager who learns to direct agents well isn't managing a team of five. They're managing a team of fifty who are available instantly, infinitely patient, and becoming more capable.

High Output Management wasn't just a book about managing people in an analog world. It turns out it was also the manual for managing agents in a digital one. We just didn't know it yet.

The 10x manager is the next frontier. And the unlock isn't learning to code. It's applying the skills you already have.

This isn't an argument against engineers. The best ones, using these tools, are operating at a level that wasn't possible two years ago. But the gap between what a great engineer can do and what a clear-thinking operator can do is closing faster than most people realize. The bottleneck is shifting from technical execution to problem definition. 

The person who can clearly articulate what needs to happen, in what order, and what good looks like has leverage that is going exponential. The engineer builds the engine. The manager decides where to drive.

If you can write a job description, you can prompt an agent.

If you can design an org chart, you can build a workflow.

If you can explain how your business works better than anyone else in the room, you have the most valuable input an agent can receive.

The engineers will figure out the infrastructure. They're already on it. 

The question is who's going to direct it, and right now, that seat is empty.

Here's a simple example of where I've found real leverage.

I send out a monthly investment idea email to my investor network. The source material is a constant stream of inbound pitch decks, one-pagers, and investor letters across funds and companies.

The old process was painfully manual. Hours of reading, summarizing, formatting, and organizing before anything useful could go out.

I rebuilt it as an agentic workflow.

Materials drop into an inbox. An agent reads each document, extracts the key information, writes a standardized summary with plain-language investor fit criteria, and assembles everything into a formatted monthly digest grouped by asset class.

A 17-deal digest, the kind of output that would have taken a junior analyst a full day, now comes out clean and consistent without me writing a line of code.

But the important part is where I still plug in.

The filtering hasn't disappeared. The judgment hasn't disappeared. I still decide what actually makes the cut. What's changed is the amount of manual work required before I can apply that judgment.

Less noise. More context. Far less time buried in decks.

Here are some things I'm exploring next.

-Alec
2/52

Watch the Karpathy interview — The Sequoia talk that started this. Worth watching even if you have to pause every five minutes to Google something.

High Output Management — Andy Grove, 1983. Still the best book ever written about getting things done through other people. Read it again with agents in mind.

Meet the Agents at USV — Union Square Ventures published a post introducing their internal team of AI agents: Arthur handles deal analysis, Sally transcribes meetings, Ellie monitors email, Connor tracks calendars. The clearest real-world example I've seen of what the 10x manager actually looks like in practice.

Your Couch-to-5K for AI — Lenny Rachitsky published this recently and it's a great practical on-ramp for non-technical people. Less than 10 minutes a day. No coding required.

Agents Over Bubbles — Ben Thompson at Stratechery making the case that agents aren't hype — they're the moment the technology actually starts working. Good context for why this matters now.

GSD Camp — My friend John Davison recently launched this for operators who want to understand what's actually happening under the hood. The goal isn't to become an engineer — it's to learn how to steer the ship. Ends with a one-day hackathon to ship something real.

Aaron Levie on X — Box CEO. Consistently the clearest plain-English voice on what AI actually means for how businesses operate.