AI Engineering
What this topic covers
The impact of AI tools and agents on software engineering practice: how workflows are changing, what new skills are required, what structural constraints determine adoption, and what the long-term shift in the craft looks like.
Core concepts
- ai-coding-spectrum — eight-level framework from no AI use to multi-agent orchestration
- dracula-effect — cognitive drain from AI-augmented work; ~3 peak hours/day at 100x throughput
- prototype-as-product — slot machine programming: generate many, ship the best
Key sources
- steve-yegge-pragmatic-engineer — Yegge on the IDE’s evolution, the AI coding spectrum, and structural blockers to enterprise adoption
Synthesis
The central claim from the current evidence base: AI does not make engineering easier — it makes it faster and harder simultaneously. The easy work disappears (automated); the hard work expands (more output requires more decisions, more review, more architectural judgment). The dracula-effect is the mechanism for this.
Two structural blockers to AI adoption are worth tracking:
-
Monolithic codebases. AI agents have an effective context ceiling (~0.5–2M LOC). Large monoliths structurally exclude teams from agentic benefits until they modularize. This maps to scaling at a higher level — the bottleneck is not compute but comprehensibility.
-
Reading ability. AI tools currently produce walls of text. Engineers who struggle with dense textual output are filtered out before they can develop AI fluency. Yegge predicts near-term interface evolution (visual avatars, audio) that lowers this barrier.
The engineering-knowledge-shifts-continuously argument is worth holding: Assembly expertise was essential in the 1990s, obsolete today. AI fluency follows the same curve. The correct response is not resistance but directed investment in the skills that remain essential after the shift.
Open edges
- What do levels 5–8 on the AI coding spectrum look like as documented workflows? No systematic treatment yet.
- How do organisations with large monoliths begin modularisation to unlock agentic workflows?
- Does the dracula-effect plateau as engineers build tolerance for sustained high-intensity work?
- What evaluation frameworks exist for the prototype-as-product model — how do you pick the “best” from 20 generated implementations reliably?
- The SaaS platform imperative (products without APIs will be outcompeted by AI-native replacements) — are there documented case studies beyond Yegge’s Zendesk example?