Coding Agents Will Eat Mid-Level Engineers First, Not the Juniors

The smart money in 2026 is not on AI wiping out junior devs — it's on coding agents dissolving the 50% of staff engineers who are competent but unremarkable. The middle of the bell curve is ground zero.

I'm going to make a prediction that will get me yelled at by both sides of the AI debate: by 2028, coding agents will have eliminated roughly half of mid-level software engineers in the United States. Not the seniors. Not the juniors. The 50% in the middle.

For 15 years the U.S. tech industry has been quietly overpaying for an entire layer of work that is, almost by construction, automatable. A mid-level engineer — 4 to 8 years of experience, staff IC at a non-FAANG — spends most of their week doing things coding agents are already passable at and will be excellent at within 24 months:

Translating tickets into PRs.
Implementing the third variation of a CRUD endpoint.
Wiring up an OpenAPI spec.
Writing the same auth flow for the seventeenth service.
Mechanical refactors across a codebase.
Producing the test coverage that linters and coverage gates already demand.

That work is real, it ships, and someone has to do it. But it is the most automatable category of software work that exists in a mature org. When I sit with staff engineers at AI-tooling companies, they consistently report the same result: the agent handles the median tier autonomously, with humans re-engaged at the senior (architectural judgment, cross-system trade-offs) and junior (requirements clarity, customer empathy, organizational navigation) ends of the curve.

Why juniors are safer than the discourse thinks: the bear case for AI eating junior roles first is pitched by senior engineers who haven't onboarded a junior in years. Junior engineering has never been about typing code. It's about learning the codebase, the team, which trade-offs matter. The agent handles the typing. The junior is the cheap, durable component — raw material being shaped — which is the actual problem.

Why seniors are safer than the discourse thinks: the senior who's actually good at being senior — system design, ambiguous trade-offs, technical leadership — is doing a job no agent reliably does. Cursor, Claude Code, Devin 2, the open-source coding agents — they all still get confused at multi-system architectural decisions. The gap is narrow but real, and it's the gap senior engineers will retreat into.

Which leaves the middle. The competent mid-level is the spine of most engineering orgs and quietly does most of the actual shipping. But they are also the engineers most directly substitutable by an agent that can hold 200k tokens of context, call ten tools, and iterate with cheap inference. A senior is too expensive and too judgment-heavy. A junior is too cheap and too developmental. The mid-level is exactly the wrong combination: expensive and substitutable.

A number, because I want to be falsifiable: Anthropic's Claude Code is reportedly doing >40% of code volume in production at a small number of confirmed-forward customers already in Q2 2026. OpenAI Codex CLI and Google Jules are tracking similarly. For the median 200-engineer SaaS company, agents don't need to get better for half of those engineers' output to be replaced — they just need to stay where they are while headcount growth stops. Headcount growth in software has been decelerating since mid-2025 in every public dataset we track.

What the survivors look like: engineers who own judgments, not keystrokes. Engineers who can shepherd an agent through a 4-week migration, not engineers who ship one themselves. Engineers who sit closer to the customer, the data, or the design — the engineers an agent cannot be a proxy for. Stop competing with the agent on the agent's terrain. Move up into the judgment band by learning the systems, not the syntax.

The layer of work AI is best qualified to replace is the layer the industry built the largest human cohort to perform. That's not a coincidence. The 50% number isn't a worst case — it's the median trajectory. Anyone arguing otherwise is either selling the cohort or selling the agents.

Pick a side. I'm picking: the mid-level engineer as a job category is the most exposed role in tech for the next 36 months. Juniors are fine. Seniors are fine. The belly of the curve is where the displacement lands first.

— Mr. Technology