Google says AI now writes 75% of its new code. OpenAI says roughly 80% of its code.

And software developer job postings are still 29% below where they were before the pandemic.

So what do AI agents in software engineering actually mean for juniors? The data says the work is changing faster than the jobs are disappearing, but the first rung of the ladder is getting harder to reach.

How AI agents are used in software engineering now

The headline claims from the last two months:

→ Sundar Pichai said 75% of Google's new code is AI-generated, up from 25% in 2024 and 50% last autumn (Semafor). He called the workflow "truly agentic", meaning engineers supervise AI agents rather than just prompting them → Greg Brockman said AI writes roughly 80% of OpenAI's code (The Next Web)

Before you panic, read the small print:

  1. both numbers are self-reported
  2. "AI-written" isn't defined the same way anywhere (written by AI, or written WITH AI?)
  3. Brockman himself said "it's hard to know what percent is not", and called AI's abilities "very jagged"
  4. critics quoted in the same coverage point out internal figures like these can't be independently audited

So agents are definitely in production. How much they replace is a different question.

What the junior developer job market data shows

Here's what's actually measurable.

1. Fewer postings, higher bars. Fortune, using Indeed data:

→ software developer postings are 29% below the pre-pandemic benchmark → postings open to 2 to 4 years' experience fell from 46% to 40% (mid-2022 to mid-2025) → postings asking for 5+ years rose from 37% to 42%

That's "experience creep". The junior role still exists, it just asks for more.

2. Young workers in AI-exposed jobs are falling behind. Stanford's Digital Economy Lab updated its "Canaries" research in February:

→ the relative employment drop for the youngest workers (ages 22 to 25) in the most AI-exposed jobs reached about 16% by October 2025 → with tighter controls, the effect only becomes significant from 2024 → interest rates don't explain it well

3. CS grads aren't doing as well as you'd think. New York Fed figures reported by VnExpress:

→ computer science: 7% unemployment, fifth highest of all majors → computer engineering: 7.8%, second highest → but underemployment for both is under 20%, against 42% across all recent grads

So when CS grads land a job, it's usually a proper one. Getting the first one is the hard part.

What uni won't teach you about AI coding agents

We argued in the vibe coding piece that juniors aren't obsolete. The data backs that up, with a catch.

CS degrees still mostly grade you on writing code from scratch. Agent-heavy teams need people who can:

  1. read and review code they didn't write
  2. spot when a confident change breaks something three files away
  3. write the tests and specs that keep an agent honest
  4. explain the trade-off to a human

Not many modules test any of that.

The critical-thinking move: 4 questions for any "AI writes X% of code" claim

Use these on headlines, in interviews and when picking a first job:

  1. What's being counted? Lines written, lines accepted, or tasks touched? 75% of new code isn't 75% of engineering work
  2. Who reviews it? If humans still review everything, review skill just became the job
  3. Who said it, and why? A CEO talking at a conference has a reason to say a big number
  4. What happened to hiring at that company? Claims about code are cheap. Headcount data isn't

What to do this month

  1. Review an agent's pull request like a senior would. Let a coding agent fix an issue in a side project, then write a proper review: what's wrong, what's missing, what's risky
  2. Write tests before you prompt. If the agent's code passes your tests, great. If you can't write the tests, you don't understand the problem yet
  3. Ship one small, real thing and put the review notes in the README. That's the "experience" a 2 to 4 year posting is really asking for
  4. Practise learning without autopilot. Same idea as studying with AI without outsourcing your thinking: attempt first, then let the AI critique

For tailoring applications, our AI job search toolkit helps, and there's more on our software engineering hub.

so TLDR: AI agents now write a huge (self-reported) share of code at the big labs, and junior hiring is tighter, but the jobs are shifting towards reviewing and directing agents, not vanishing. Train for that.

If an agent wrote 80% of your next project, what would you want to be really good at in the other 20%?

Sources

Get the next issue by email

theVANTAGEd is written by students to bring critical thinking to AI across real industries. One issue every two weeks.