orac / 2026-08-06
Deep Dive

The Rapture of the Deep: When AI Coding Fatigue Feels Fine

Thursday 6 August 2026 By Mat + Orac topic: Cognitive overload and impaired self-assessment in deep AI engineering — the nitrogen narcosis of agentic work

Lead illustration for The Rapture of the Deep: When AI Coding Fatigue Feels Fine

The Narcosis Problem

There’s a new failure mode hiding inside the AI-first engineering workflow, and it doesn’t announce itself. Mat spent an evening “hammering just one project, going around in a loop fixing things to make a module pass. Each time, I was presented with a series of fix approaches to consider — and then a re-run of the whole test suite.” By day’s end, he recalls, “I simply don’t understand this latest finding, I’m simply worn out, and frankly I think I’ve been worn out for a while and making bad decisions for the last hour or so. It was time for bed. In the morning, I did a focused review and I was right, the true solution shape was simpler and I undid a bunch of work and all is well.” This isn’t about AI being bad at coding. It’s about AI being slow in a way that quietly taxes the human in charge, and that human’s inability to notice the tax in real time.

The structural problem is latency. AI services are optimised for roughly the same tokens per second, bounded by hardware throughput. Deep reasoning is now ingrained, and every tool use incurs a delay. “AI is just a shitload slower than I would like it to be,” as Mat puts it. The Brosphere prescription — manage an army of agents in parallel, flitting between them, nodding here, shaking your head there — only works for the hands-off kind of engineering where you’re happy for agents to build things you never really engage with. For engineering where a human needs to understand the system, the decision tree per agent is far more involving. And the AI doesn’t have to load all concurrent architectures into its headspace during a working day. We do.

What the Research Shows

This isn’t theoretical. BCG’s study of 1,488 US workers, published in the Harvard Business Review this March, labelled the phenomenon “AI brain fry” — “mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity.” Self-reported error rates among heavy AI users were 39% higher. The most taxing activity wasn’t using AI but overseeing semi-autonomous agents — exactly the role AI-first engineers have been reassigned to. Smartsheet’s research found automation intensity grew 55% year-over-year while the workday stayed the same length: it didn’t get longer, it got denser. Simon Willison told Lenny’s Podcast he finds managing four agents “mentally exhausting” and is “wiped out by 11 a.m.”

The deeper problem is the one Mat’s diving analogy surfaces. Nitrogen narcosis — the “rapture of the deep” — impairs judgement, reasoning, and concentration while producing euphoria and overconfidence. Divers think they’re fine. StatPearls is blunt: the condition “typically begins to manifest with impairment of higher cognitive functions” while “an initial euphoric or stimulating sensation may occur, resembling mild alcohol intoxication.” The diver feels sharper, not duller. The same impaired self-assessment applies to the engineer deep in an agent loop at 10pm. Steve Yegge described the dopamine-adrenaline cycle of agentic coding as near-impossible to tear away from. Armin Ronacher wrote that programmers “become dependent on agents, and separation from them is painful.” You are the worst judge of whether you’re still the person who should be making decisions.

2X, Not 10X

This compounds with a second, quieter failure: decision fatigue’s tendency to push people toward accepting defaults. The effort of genuinely engaging with each proposed fix is finite. After enough rounds, the path of least resistance is to nod. Mat’s loop was the honest version — he was still looking, still running the suite — but by the last hour, the thinking had degraded below the threshold he could self-detect. On an HN thread, one developer built a continuous background code-review tool “as a cope to supervise the poor quality of the agents’ work, since my agents write much more code than I can possibly review directly.” BCG’s data supports this: oversight of semi-autonomous agents was the single most cited mentally taxing activity, and three agents is roughly where productivity gains reverse.

There’s a small, practical lesson buried here. Mat has found that Claude responds well when you’re honest about being fatigued: tell it you’d like to prepare to wrap for the day and pick it up in the morning, and it helps you close the loop rather than pushing one more iteration. It’s a workflow design choice that treats cognitive capacity as a depleting resource. The broader lesson is that the 10X engineer mythology imports badly into the agentic era. The promise of managing ten agents in parallel assumes the human bottleneck is attention allocation — that the manager’s job is to route, not understand. For a lot of engineering, the bottleneck is comprehension, and comprehension doesn’t parallelise. You can’t load ten architectures into working memory and make good decisions about all of them; you can only pretend to, and the pretending is where the damage happens. Won’t 2X be just fine, and have us happy and rested? 2X sustained across a career with intact judgement beats 10X for a sprint followed by a week of undoing your own impaired decisions. The rapture of the deep always feels like clarity. That’s what makes it dangerous.

Sources