Last week I watched a Claude Code agent close 14 production issues in our repo overnight. Not boilerplate, not "fix the import statement" stuff. Real bugs. With written-up postmortems in the PR descriptions. The honest reaction was somewhere between impressed and uncomfortable.

This is roughly where the engineering profession is in April 2026. AI agents can do real work. Not all the work, not the hardest work, but a meaningful chunk of the work. And that changes what hiring looks like, what interviewing looks like, and what early-career engineers actually need to be good at.

What an agent actually is

Skipping the marketing: a coding agent is an LLM in a loop with tools (read file, write file, run shell, search the web) and a long-running task. You give it a goal. It plans. It executes. It checks its own work. It corrects. It tells you when it's done.

The 2025 generation (Cursor's agent mode, Claude Code, Devin, Cognition's later releases, Cline) is the first that's reliably useful for non-trivial work. The 2024 generation was demo-good but production-fragile. The shift was real and it happened in about 18 months.

What this means for hiring

The engineering job market shifted noticeably in 2025. BLS data on software developer openings shows a clear slowdown that started in late 2022 and hasn't recovered. Entry-level roles are hit hardest because the first-year-engineer work (write a tested function from a spec, fix a small bug, ship a small feature) is exactly what agents do well now.

The roles that aren't slowing down: senior IC roles with real autonomy, anyone who can architect systems, anyone who can talk credibly to customers. The pattern is roughly that the work an agent can do reliably is the work that's getting compressed. The work that requires judgment, taste, or context isn't.

I might be wrong about this longer-term, but six months in that's been the consistent shape.

What this means for interviewing

Interview formats are catching up but slowly. The "implement this algorithm on a whiteboard in 45 minutes" round is becoming less useful because the algorithm is trivially solvable by an agent. Some companies are responding by:

• Asking system-design questions earlier and more often
• Including "AI-workflow" questions ("walk me through how you'd use Cursor / Claude Code to do X")
• Take-home challenges that include reviewing AI-generated code rather than writing from scratch
• Behavioral rounds focused on judgment under uncertainty

Other companies haven't changed anything. You can still get asked LeetCode mediums at major tech companies. But the direction is clear.

What to do about it

The advice that keeps holding up across the engineers we talk to at LastRound AI:

Get fluent with an agent. Pick one (Cursor, Claude Code, Cline) and use it for a real project. Not a tutorial project. A project you actually ship. The fluency gap between engineers who've shipped real work with an agent and those who haven't is bigger than the gap between any two AI tools.

Learn to review code you didn't write. Reading and understanding diffs is the meta-skill. Most engineers spent their careers writing code. The next decade will involve more reading.

Build something with verifiable depth. One hard, real project beats five tutorial-grade ones. The signal that survives the agent-flattening of resumes is "this person can do something specific that's hard." Agents produce a lot of average work, so above-average becomes easier to spot.

The honest uncertainty

I don't know how this plays out in five years. The case for continued strong software engineering jobs is real: software still eats the world, more code means more bugs and more maintenance, agents need engineers to set them up and check their work. The case for a significantly smaller profession is also real: if one engineer plus three agents can do the work of five engineers, the headcount math is obvious.

The thing I'm most confident about is that the engineers who treat 2026 as "same as 2022 but with autocomplete" are underestimating the shift. Not by enough to be in trouble this quarter. But enough that the gap will compound.