Five Notes from the Sequoia Stage
Andrej Karpathy spent thirty mostly technical minutes on the Sequoia stage describing the executive shift of the decade. Five things from it survive translation.
The autopilot does not absolve the captain; it promotes the captain. Once the machine flies the airplane, the human is no longer paid to push pedals. The human is paid to know which pedal the machine should not be allowed to push. That is the small inversion at the centre of Andrej Karpathy’s From Vibe Coding to Agentic Engineering, given on Sequoia Capital’s channel this week, and it deserves a translation. He spent thirty mostly technical minutes describing the executive shift of the decade. Five things from it survive translation, and each one is a budget conversation wearing a coat.
1. December 2025 was the inflection. Reprice the memory.
The most expensive item on most balance sheets is not a server. It is a memory of a 2024 demo. We are pricing 2026 decisions in 2024 dollars of belief, which is a charming way to lose money slowly. The gap between what the technology was eighteen months ago and what it is today is wider than the gap between today’s technology and a competent junior engineer, and the junior engineer at least has the manners to take a weekend.
The agentic loop simply began to work. Not better in the patient way of things that improve with caveats, but better in the rarer way of things that cross a line and decline to come back. The agent reads the repository, writes the change, runs the tests, fixes its own break, opens the PR, and waits for someone wiser to look at it. We have spent considerable capital insuring against a future that has had the indelicacy to arrive early. Reprice it. The vendors we were right to doubt last year are not the vendors to doubt now.
2. Software 3.0: prose is the new program.
In the 1950s, Karpathy reminds us, it was not obvious whether computers would end up looking like calculators or like neural nets. We picked the calculator. For seventy years we have been writing instructions in the calculator’s grammar: precise, deterministic, brittle, alien. Every shell script, every API contract, every regex of doom is a small monument to the paradigm we chose.
The neural net never went away. It has been running, all this time, virtualised on top of the calculator. What changed in the last eighteen months is that the neural net has become competent enough to climb back up the stack. It is slowly becoming the host process, and the calculator is being quietly demoted to a co-processor for the deterministic work the model still finds tedious.
The visible consequence is that the instruction set has begun migrating from the machine back to the mother tongue. Where an install used to be a shell script, it is now a paragraph of English that an agent reads and obeys. The compiler is the model, and the model is, on its better days, a literalist with a working memory.
Read the implication twice, because we usually read it once and then file it. The runbooks, the specs, the README nobody updates, the half-loved wiki page about how the billing service handles refunds: all of it is now runtime. The polite request we wrote for the next human is now an instruction the agent will execute, faithfully and at speed, including the parts we got wrong on a Friday.
Documentation has quietly become load-bearing engineering, which is a sentence the technical writers have been waiting twenty years to hear out loud. Fund them. Refactor the wiki. The companies that treat docs as overhead in 2026 will, by 2027, be the companies wondering aloud why their agents underperform a competitor’s agents on the same model.
3. Intelligence is jagged. Govern locally.
The same model that refactors a hundred thousand lines of legacy code will, in the very next breath, advise us to walk across six lanes of traffic to reach a fifty-metre car wash. Brilliance and absurdity rent the same apartment. The capability surface is jagged, and pretending otherwise is the most expensive form of optimism currently on offer.
This breaks the global trust dial that most governance frameworks were quietly built around. There is no setting marked “trust the model 73 percent.” There is only this: in this domain, with this scaffolding, with this reviewer, the model performs at this level, and outside that envelope all bets are off and most of them are wrong.
The operational consequence is that risk frameworks must be local. Per workflow. Per team. Per blast radius. The CISO who promises one org-wide AI policy is promising something the technology cannot underwrite, and the technology, unlike the CISO, will not be polite about it. Push governance to where the work happens, where the people using the tool can tell the difference between brilliance and a suggested stroll into traffic.
4. Fund the floor. Defend the ceiling.
Vibe coding raised the floor. Anyone with intent can ship something now, and we should not pretend this is a tragedy. Product managers prototype their own ideas. Sales engineers build real demos. Finance writes its own dashboards and, occasionally, its own opinions. The democratisation is real, and it is welcome, and it would be ungenerous to grumble.
Agentic engineering preserves the ceiling. Specs, tests, review, taste, the unglamorous habits that keep a system from collapsing under the weight of its own first draft. These are what hold the ceiling at a professional altitude rather than at the altitude of a particularly enthusiastic intern.
The strategic error of the next eighteen months will be conflating the two and quietly defunding the discipline. On the spreadsheet it will look like efficiency. In production, six quarters later, it will look like a slow leak that nobody can quite locate. Anyone can build a feature and anyone can build a system are sentences with the same grammar and very different consequences. The institutions that hear the difference will fund both. The ones that do not will discover, expensively, which one they had been paying for all along.
5. We can outsource the thinking. We cannot outsource the understanding.
Karpathy closed on this, and so will we. The temptation, particularly at the executive layer, is to hear it as a comment about cognition. It is not. It is a comment about ownership, which is a quieter and more demanding word.
The signature on the decision is ours. The understanding is what the signature signs for. We can delegate the drafting, the analysis, the synthesis, the slide that nobody enjoyed making. The irreducible part of the executive job is to know what we are putting our name to. Without that, we are not delegating; we are abdicating, and an organisation can smell abdication within a quarter and price it within two.
This is the part of the talk that belongs in board packets. Not as a warning, but as a job description that finally fits.
The competence to wield
Five highlights. One phrase beneath them all: competence to wield. The agentic loop works now. Prose is executable. The intelligence is jagged. The discipline must be funded. The signature stays with the human, where it has always quietly belonged.
The institutions that build this competence will look, in 2027, as if they got lucky. They will not have gotten lucky. They will have funded the floor, defended the ceiling, and kept their signatures honest while everyone else was busy keeping them busy. The rest will be running 2024 playbooks against 2026 competitors and wondering, with genuine puzzlement, where the margin went.
We have the tools. We owe them the practice.
The autopilot is on. The captaincy is ours. Fly the airplane.
