“Code without juniors”

What if AI coding agents write most new code, while large companies and offshore IT centers cut graduate hiring?

Companies almost stop hiring junior developers to write code: the bottleneck has shifted from writing code to reviewing it, and review is entrusted to experienced people. People enter the profession through reviewing and maintaining AI code. A few years later a gap appears: few have had time to go from junior to experienced engineer.

Which technologies this leads to

• Agentic engineering+AI pair programmer → AI developer mentor^*
  The first job, where a junior learned from other people's code, is replaced by a mentor agent: it assigns review tasks, goes over mistakes, explains how the system works.
• Autonomous AI agents+Large language model → AI code audit
  Review becomes the main work and the main bottleneck, so tools are needed that check AI code faster and more deeply than people.
• Generative AI+Open source → Work-experience simulator^*
  Attempts are made to compress the experience that years of work used to give: synthetic projects with real bugs from open source, used for training instead of an internship.
• Autonomous AI agents+Large language model → Codebase memory^*
  Fewer and fewer people remember why a system is built the way it is, so this knowledge is moved into the project's machine memory, which both people and agents read.

These are potential technologies: working names (*), no articles yet. In the technology tree they are shown with a dashed arrow ⇢ and the scenario number.

Consequences

• Work & jobs: Entry into the profession is through reviewing and maintaining AI code; internships in writing code become scarce.
• Education: Universities and courses shift their teaching from writing code to review, architecture and assigning tasks to agents.

• Software development: A shortage of experienced engineers and rising salaries for them; knowledge of systems rests on fewer and fewer people.
• Cybersecurity: Fewer people understand code in full, so the risk of unnoticed vulnerabilities grows (see SC-2026-0002).

Grounds

• At Google, 75% of new code is generated by AI and approved by engineers; in autumn 2025 it was 50%. — blog.google, 22.04.2026 ↗ · a claim by a company that sells AI tools
• On the SWE-bench Pro benchmark, the best models solve about 60% of real code-fixing tasks; in 2025 it was about 23%. — Scale AI, SWE-bench Pro leaderboard, 2026 ↗ · the table is undated
• 66% of developers who use AI work with AI agents, 73% of them daily. — Stack Overflow Developer Survey, 06.10.2026 ↗ · a voluntary survey; shares are among AI users
• Vibe coding+Autonomous AI agents → Agentic engineering (2026)

• Work & jobs / Tobi Lütke: “Teams must demonstrate why they cannot get what they want done using AI”

• Employment of 22–25-year-olds in the occupations most exposed to AI is 19% lower than if it had grown like that in less exposed ones; experienced workers show no such gap. — Stanford, “Canaries in the Coal Mine?”, 12.08.2026 ↗ · ADP payroll data; the link is descriptive, not causal
• Graduate hiring at the 12 largest tech companies fell about 65% compared with 2019, and at early-stage startups by 76%. — SignalFire, State of Tech Talent 2026, 22.06.2026 ↗ · a venture fund; the effect of AI is not separated from the end of the cheap-money era
• Of US software development job postings, 69.3% are senior level and 4.5% entry level (Q1 2026). — Indeed Hiring Lab, 23.07.2026 ↗ · the posting level is set by Indeed's classifier
• At companies that adopted generative AI, junior employment falls because hiring slows; senior employment does not change (65 million résumés, 280,000 companies). — Hosseini, Lichtinger (Harvard), 06.05.2026 ↗ · working paper, not peer-reviewed

• TCS hired more than 44,000 graduates in fiscal 2025/26 and made 25,000 offers for 2026/27. — Business Today, 13.04.2026 ↗ · the company ties this to demand, not directly to AI
• Infosys keeps graduate hiring at 20,000 in 2026/27, as a year earlier; headcount fell by 8,440 over the quarter. — Outlook Business, 24.04.2026 ↗
• The $100,000 fee for US H-1B visas (since 19.09.2025) was extended to 2027 but struck down by a federal court on 08.06.2026; the appeals court did not stay that ruling. — Manifest Law, 19.09.2026 ↗ · a law firm's blog; the effect on offshoring is not measured
• Unemployment among 16–24-year-olds in China was 18.9% in August 2026, with a record 12.7 million graduates. — Trading Economics, based on data from China's National Bureau of Statistics, 17.09.2026 ↗ ⧗ · a secondary source; the link to AI is the aggregator's assessment

• Data center electricity use grew 17% in 2025 and will double by 2030; grids, turbines and transformers limit how fast they can be built. — IEA, 16.04.2026 ↗
• The price of reaching a fixed level of model capability falls on average about 50-fold per year (from 9- to 900-fold, depending on the benchmark). — Epoch AI, 12.03.2025 ↗ · Epoch warns that the fastest declines may not continue

Energy limits how fast data centers are built, but not the price of a task: no evidence was found that this slows the replacement of juniors.

No data: no regulation directly concerning the hiring of junior developers was found.

• According to Y Combinator, a quarter of the startups in its winter batch have 95% of their code written by AI · 06 Mar 2025
• METR: experienced developers with AI worked 19% slower, though they believed they had sped up · 10 Jul 2025

Mechanism

1. AI writes more and more code: 75% of new code at Google; agents are used daily.
2. The bottleneck shifts from writing code to reviewing it: “junior engineers can slow a team down” (SignalFire, 2026).
3. Companies that adopted AI hire juniors more slowly: 4.5% of postings are entry level, graduate hiring is −65% compared with 2019.
4. Offshore centers are cutting hiring too: TCS went from 44,000 to 25,000 offers.
5. A few years later, few have made it to experienced engineer, and a gap appears.

Against

• ⚑ Matt Garman: “With no talent pipeline, the whole thing explodes on itself”
• Across the economy as a whole, there is no clear trace of AI in the labor market: occupational change follows pre-AI trends. — Yale Budget Lab, 15.09.2026 ↗ · a snapshot with no breakdown by age or occupation
• US software development postings are 15% higher than at the launch of Claude Code (02.2025), although total postings are 7% lower; but 71% of the growth is senior positions. — Indeed Hiring Lab, 08.07.2026 ↗ · correlation, preliminary data
• The US Bureau of Labor Statistics projects 10% growth in the number of developers over 2025–2035 (the previous projection was 15.8%). — BLS, 27.08.2026 ↗
• IBM is tripling US entry-level hiring in 2026, including developers, but is changing their work toward clients and AI oversight. — Fortune, 13.02.2026 ↗ ⧗ · no absolute numbers
• The 2019–2022 baseline was a period of overheated hiring. SignalFire partly attributes the decline to the end of the cheap-money era; Stanford tested this explanation and found no strong support for it. — Stanford, 09.02.2026 ↗

Signposts

If these indicators keep moving the same way, the scenario is getting closer. The pipeline tracks them and brings new signals.

Affects

Junior developer · Agentic engineering · AI code audit · Work & jobs · Education · Software development