keep the human in the loop

For a couple of years now, the hype around “Agentic AI” has been completely out of control. It hits hardest inside software development, a field where improvised “Agentic Engineering” is putting on an unprecedented shit-show.

Consider what’s actually being sold. Cognition introduced Devin as “the first AI software engineer”. Jensen Huang went further, declaring that every company’s IT department is destined to become “the HR department of AI agents”. Dario Amodei went further still by saying AI “could wipe out half of all entry-level white-collar jobs” in the next one to five years. One hire, then a whole workforce, then half of it gone. Each announced proudly, by the people profiting from it.

Notice what the words are doing: human titles and corporate vocabulary draped over what is, mechanically, models calling tools in a loop until the output looks plausible. The loop can be genuinely useful — I use it — but the vocabulary never described the mechanism. These words aren’t neutral: call it an assistant and you review it, call it an engineer and you trust it.

The real problem isn’t the marketing though, it’s the workflow it legitimizes. If the thing is an engineer, why review its diffs? If it has an HR department, why keep a human in the loop? So teams fire-and-forget, merging unread code at a pace no human could audit, and quality quietly becomes the one metric nobody puts on the dashboard.

You don’t have to take my word for the outcome, it’s already in production. Satya Nadella claims that 20 to 30 percent of Microsoft’s code is now written by AI, and the results are on display:

• the Recall feature pulled before launch for being a security hole with a UI
• a Windows update cadence that racked up twenty-plus major disasters in 2025 alone
• a GitHub status page that has become Microsoft’s most reliable product

As Mario Zechner puts it in his excellent post on slowing down: “memory leaks in the gigabytes, UI glitches, broken-ass features, crashes: that is not the seal of quality they think it is”.

The panic about AI “replacing” programmers rests on one conflation: if the job is typing, a typing machine is a hire. But typing never was the job. The job was deciding what to type and knowing why. So outsourcing the typing, by itself, costs you nothing; outsourcing the thinking costs you everything. And unlike typing, thinking doesn’t come with an audit trail, nobody can diff your judgment.

What makes it worse is that these tools are engineered to be addictive. Every prompt that comes back with something that looks finished is a small reward for skipping the hard parts: the friction of the blank file, of the design you have to hold in your head, of the bug you only understand by writing it wrong first. As Armin Ronacher said: “without friction you can’t steer”. Teams run on the same friction — the reviews, the arguments that grind a design into shape. Discard it for speed, as companies now do, and you get a worse product built by an engineer who dulls with every skipped rep.

These tools give us something genuinely rare: the chance to improve speed and quality at the same time. Trading one for the other isn’t pragmatism, it’s picking the tool that makes things effortless over the one that makes the work good.

So keep the human in the loop, whatever your loop looks like. These tools are steered in natural language, so there is no playbook: the workflow that keeps you in control is personal. What isn’t personal is the principle: The moment you stop directing and start accepting, the tool isn’t working for you anymore, you’re supervising its output on its terms. Staying in the loop isn’t overhead, it’s the job. Keep the friction, on purpose. It’s what keeps you sane.

The hype will move on. Your judgment stays.