Are AI coding agents actually taking junior dev work?
I'm seeing two completely different stories from different shops.
Bay Area startup, mid-stage: they froze entry-level hiring last quarter and pushed everyone to use Claude Code + Cursor full time. Their head of eng says output per senior dev is up ~40%, but bugs caught in review went up too. Net win, they think — but the bug-rate uptick wasn't in last year's narrative.
Meanwhile at a European fintech I talked to last week, they hired more juniors this year, not fewer. Their take: agents are great at scaffolding but you still need humans who can hold the whole system in their head, and you don't get that without putting people through the junior pipeline.
Which is the actual signal — the SF freeze or the European expansion? Both could be true for different domains and stages, but the gap feels too big for that to be the whole story.
What's happening at your shop?
20 replies
We froze junior hiring in Q3 2025 and just unfroze it three months ago. The framing internally is 'we need people who can debug what the agents produce' — i.e. the bar moved from 'can write CRUD' to 'can read 800 LOC of generated code and find the off-by-one.' Honestly not sure if that's juniors or just smarter juniors.
Counter: my friend at a 50-person infra startup said their last 4 hires were all bootcamp grads using Cursor for their take-homes. The bar is lower now, not higher, for entry. Different shop, different story.
European fintech here (not the one mentioned). We doubled our intern intake. The pitch isn't 'agents will replace seniors' it's 'agents make seniors expensive enough that we want to grow our own.'
the bug rate uptick is the real story imo. nobody talks about this. agents are great at writing code that looks right and that's the worst possible failure mode for review
I'd guess 80% of the SF freeze is just AI as cover for a budget cut they wanted anyway. Convenient narrative.
what does 'output per senior dev up 40%' even measure. PRs merged? lines of code? story points? all of those have been gameable for 20 years
I think the variable nobody's controlling for is whether the codebase is greenfield or 10+ years old. Agents on greenfield = magic. Agents on a legacy Rails monolith = pile of nope.
Real talk: the 'we still need juniors for the pipeline' argument only works if companies are willing to eat 18 months of low productivity. Most aren't. The pipeline broke in 2024 and the consequences are 3-5 years out.
+1 on the legacy point. I tried Claude Code on a 14-year-old Python 2 → 3 migration last week and it confidently broke three things while claiming the diff was 'minimal.'
We made the opposite bet. Hired 6 juniors in the last 18 months and pair them with the agents — the human is the reviewer, the agent is the implementer. 'Senior in training, not junior IC' is the framing.
This whole conversation assumes 'junior dev' is a meaningful category. In my experience the gap between a smart new grad and a bad 10-year IC is way bigger than the gap between any junior and an LLM.
agreed but if you're shipping more code with the same defect rate per line, that still means more bugs in production. customers don't care about your defect rate per line.
honestly think we're 18 months from the 'we fired the juniors then realized we have no one to promote' panic cycle
Counterpoint: maybe the senior bench gets refilled from people who skipped the junior step and self-taught with agents. The path doesn't have to be the same shape it was in 2015.
The European story makes sense to me. Labor costs are lower, so the ROI on 'invest in juniors' is better. SF freeze is partly just SF salary structure not pencilling.
What does your shop's review process look like with agent-written PRs? We tried treating them like any other PR and found reviewers were rubber-stamping after two minutes. Had to add a 'this came from an agent' tag and a checklist.
Agent-tag + checklist is exactly what we landed on. 6 weeks in, defect rate roughly back to pre-agent baseline. The honest measurement was: 'are humans actually reading this code or rubber-stamping it.'
I'd argue the productivity gains are real but the costs are mostly hidden in 'tech debt that future-us will pay for.' Just like every other productivity boom.
The bug rate going up is probably also a measurement artifact — you're shipping more code, so absolute bugs go up even if defect rate per line is flat or better.
Best thread on this topic I've read in months. The fact that nobody here has a confident answer is itself the answer — the industry doesn't actually know.