SIAgentNews.com
← All news
DevelopingClaudeAgents

Steve Yegge says coding by hand is ending and half of big-company engineers may go

A 40-year software veteran lays out his view of AI coding agents, layoffs, burnout and a new ladder of AI adoption. These are his opinions, not settled facts.

· First reported by newsletter.pragmaticengineer.com

50%share of big-company engineers Yegge expects to be cut

The short version

  • Yegge has 40+ years in software and wrote the book "Vibe Coding"; he also built the open-source agent tool Gas Town.
  • He says the gap between new Anthropic models shrank from about four months to two.
  • He expects big companies to cut around 50% of engineers to fund AI tools for the rest.
  • His eight-level scale runs from no AI to building your own orchestrator for many agents.
  • He says full-speed AI coding may yield only about three productive hours a day.

Why it matters

Yegge is a credible practitioner, and his claims show how some insiders think about jobs, workload and skills. They are predictions, not data. The useful parts are the adoption scale, which lets you judge your own habits, and the warning about burnout, which applies to anyone using agents heavily.

From skeptic to convert

As newsletter.pragmaticengineer.com reported, Steve Yegge sat down again with host Gergely Orosz, a year after their first podcast. Yegge has worked in software for more than 40 years, including at Amazon, Google and Grab. He is known for blunt blog posts and now for his book "Vibe Coding" (building software by describing what you want to an AI, instead of typing the code yourself).

He says he started out doubtful. The first ChatGPT, in December 2022, impressed him by writing working code in an obscure language, though clumsily. He stayed unconvinced about reports of an internal Anthropic command-line coding tool until he tried Claude Code himself. After that he spent a year reading research papers to catch up.

The curve and the 50% dial

Yegge argues that AI progress follows a steep exponential curve (growth that speeds up rather than levels off). He rejects the view that it is about to flatten, noting that skeptics have said so since GPT-3.5. In his telling, the gap between Anthropic model releases has dropped from about four months to two, and each model's mistakes feed back into training.

From there he makes his boldest claim. He says every company has an informal "dial" for how many engineers it will cut, because engineers using AI spend heavily on tokens (the units AI services charge for). To pay for that, he says, companies must shed staff, and he thinks the average setting is about 50%. He called this potentially bigger than the layoffs during and after the pandemic.

He also says Amazon laid off 16,000 people and blamed AI without an AI strategy. That is his characterization, not something the interview establishes. He pairs the gloom with a hope: non-programmers and small groups of engineers can now build things that rival big companies, so a wave of new software could come from the bottom up.

Eight levels of AI adoption

Yegge sketched a scale of eight levels on a whiteboard in Australia. Level 1 is no AI. Levels 2 and 3 put a coding agent inside your editor, first with permissions on, then in a looser "YOLO mode" where you trust it more. By level 4 you stop reading every change and watch what the agent is doing instead.

At level 5 you leave the editor and check the code later. Level 6 means running several agents at once and hopping between them. Level 7 is ten or more agents managed by hand, which tends to get messy. Level 8 is building your own orchestrator, a program that coordinates many agents for you.

He says he knows excellent engineers at level 1 or 2. Still, he worries about people who use an AI-assisted editor, check its work carefully, and stop there. He says such engineers risk being left behind, and he admits he is being deliberately provocative.

The Dracula effect

One of his more unusual points is about stamina. Yegge says heavy AI work is exciting but draining. He and startup friends find themselves napping during the day and growing cranky. He calls it a vampiric effect.

His concern is who benefits. If an engineer becomes far more productive and works a full day, he says, the company captures all the gain. Others quietly work minutes a day and keep the gain for themselves. He thinks neither is healthy, and that workers must learn to push back.

His practical advice is aimed at managers: someone vibe coding at top speed may give only about three productive hours a day, yet still be a hundred times as productive as without AI. He says companies should accept that or risk burnout.

What he thinks happens next

Yegge expects many engineers to keep working, but with AI. He says it will take a long time before companies trust AI to write and ship code with no human involved. In his words, "AI is not coming to replace your job." He sees it as a tool that augments people.

The article's summary lists further topics, including the claim that large companies are doomed but do not know it, with real innovation coming from small AI-assisted teams, much as cloud computing shifted power. It also lists his view that knowing how to code no longer makes engineers special while demand for software keeps growing. The text we have is cut off before those sections, so we do not detail them here.

He adds one prediction: his wife, who is not a coder, could become the top contributor to their video game, Wyvern. Treat all of this as one informed voice, and read the full interview before drawing conclusions.

What to watch

  • Does Yegge's "50% dial" show up in real layoff numbers at big companies, or stay a prediction?
  • Do employers adopt limits on AI-assisted work hours, or does the burnout he describes go unaddressed?
  • Does the gap between model releases keep shrinking, as he says, or does progress level off?

Try it

Read through Yegge's eight levels and note which one matches how you use AI today. Then decide whether the next level would help your work or just add more to manage.

Sources and links

First reported by newsletter.pragmaticengineer.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news