AI-Native: Understanding Can't Be Delegated

Litvinov introduces himself: over thirteen years in development, a principal engineer by big-tech standards, years spent around ML and then LLMs; there was a management stretch too — a department of more than a hundred engineers and engineering strategy at director level. His first large-scale use came at Lyft, where he worked on ride safety: a billion rides a year, and he tried models on pull requests and his own development; he began pasting into a chat back in 2022, on GPT-3.5, and believed in it strongly. The book grew out of notes in Notion: he was tired of repeating himself like a parrot at meetups and training sessions, and no assembled system existed in English, let alone Russian. He planned a hundred, maybe a hundred and fifty, pages and ended up, by his own account, with around 650: lessons, reusable patterns and a comparison of existing methodologies. Alexander's feedback called it difficult and offered a metaphor: there is a manual for the driver and a book for the mechanic who rebuilds the engine — this is the second kind.

The frame of the conversation is a nine-level ladder; Alexander recalls it as levels of maturity with models, though he cannot remember the exact name. From the bottom up: writing code yourself, free chats, paid chats, a CLI agent in the repository with endless approvals, full access, parallel agents in separate terminals, a team of agents with roles, memory and shared synchronisation points, then orchestration and an ecosystem. The ladder was not invented but copied from how people actually move. The barrier at each step is psychological: last week someone told Litvinov he had pressed the confirmation button some hundred and fifty times and thought he had worn it out. Alexander adds that it is not only about trust: security may allow the agent yet forbid the mode without confirmations — and people, he says, train well and soon get used to pressing the button; Litvinov picks it up: attention blurs and you tap it like the hamster in that recent trend. Litvinov sees no mature orchestrators yet, and on today's models such a setup costs tens of thousands of dollars.

The top, ecosystem level is governed, he argues, only by principles — semantic invariants explained through the domain language of DDD: an expert states the task in a way that lands in the right zones of the model's semantic space, and a lawyer you know will get a better answer first time than you will, however good your prompting. Alexander picks this up: the June study from Anthropic on the persistent returns to expertise points the same way — the deeper the expertise in the task, the more successful the session and the more steps the agent takes before a human intervenes; though both classifiers were trained on the same data and may be collinear. Litvinov admits some of the book's claims are his own knowledge without proof. To switch on full access safely he separates engineering tasks from creative ones, defines the minimum expected change and the forbidden actions, and puts the boundaries outside: permissions, the shell, a sandbox and Git history cap the damage even when the agent believes it is free.

Verification debt is the debt of checking and understanding what has changed. A thousand-line pull request can now be produced by anyone, a junior included, while working through it properly takes an hour or more; it all lands on the senior standing as a living shield between the system and the bugs. The problem sits in the shadows: those accountable spend their time peering at other people's code instead of applying domain expertise, so Litvinov's first move with a CTO is to measure how long a task sits in the review column — not every company tracks time per column. The cure is not reading the diff but a package of evidence: a behavioural, usually visual change. The simplest example is a YouTube button that was red and is now blue — two screenshots, on iPhone, on Android, on a tablet. For now it is done clumsily on top of Git, some enterprises still run SVN or Mercurial, and in a heavily regulated domain the reviewer is an accountant or a lawyer.

Next comes the loss of understanding: you did not verify, you no longer read the code, you give up cognitively and let the loop spin until it has burned, say, ten thousand dollars with the system still broken. Hence Litvinov's law: what is done and how it is done can be delegated, but understanding cannot. Multi-agent work he calls a time machine — it accelerates a good trajectory and decay alike, so the loop must first close confidently on a single agent: a goal, exit criteria, feedforward as rules and architectural descriptions, feedback as hooks and tests before a commit — mechanisms that live outside the agent. He considers the ultra modes in Claude and Codex dangerous: they fail to help in 99% of cases, and someone who would have wiped his database within a year wipes it within a month once twelve times as many agents are running. Nor can you copy another company's organisation of agents — just as you cannot repeat Uber without building a company of sixty thousand people.

The final third is about people, and Litvinov has two complaints: layoffs and interviews. You hired because there were not enough hands for the products — so take on the endless backlog, otherwise the business degrades. Alexander qualifies this: the effect is spread unevenly. The frontier labs talk about replacing developers and keep hiring, because they know where to put the extra throughput, whereas in a product that is a cash cow with diminishing returns on features, or dying with five years of backlog invented by a product manager, the rollout reads as cost cutting. At the bank where he works, a team built in a week a service that uses transactions to show where you can fill up with the fuel you need, shipped it to employees and, in the second week, to users, reaching by his account some 400,000 daily active users. Both agree that the interview needs to change too: instead of drawing boxes and questions about BigDecimal, live work on a brownfield task alongside an agent. The episode closes with an announcement of Friday's conversation about agents that work with users.