A pattern keeps surfacing in developer discussions in 2026, often phrased almost identically regardless of who's telling it: a project manager takes a spec straight to an AI coding agent, ships the output, and the developer who used to own that work gets pulled in mainly when something breaks in a way the PM can't untangle. Nobody's announcing "developers are being phased out" — but the day-to-day shape of the job has shifted enough on some teams that people are asking the question openly.
What's actually driving this
It's not that PMs suddenly learned to code. It's that a meaningful chunk of what used to require a developer — translating a clear spec into working code for well-defined, low-ambiguity tasks — is exactly the kind of task current AI agents handle reasonably well without deep engineering judgment. The skill that's becoming scarce isn't "can write code," it's "can catch what the agent got subtly wrong," and that skill doesn't obviously belong to the person closest to the keyboard by default anymore.
Research on AI-assisted development gives some shape to why this is risky as a long-term pattern, not just an efficiency win. A randomized trial from Anthropic found developers using AI assistance scored notably lower on a comprehension test afterward than developers who worked unaided — with the steepest gap specifically in debugging, the skill most relevant to "catching what an agent got wrong." If that skill is what's supposed to keep this new workflow safe, and it's exactly the skill AI-assisted work tends to erode, the workflow may be less stable than it currently feels.
Is this universal, or specific to certain teams?
Worth being honest about: this pattern gets reported anecdotally far more than it's been measured systematically. It's plausible some studios and teams have leaned into this shift aggressively while most haven't changed much at all — the loud, visible cases aren't necessarily the median case. What does seem consistent across the reports is the direction, not necessarily the speed: review and verification work is growing as a share of the job, even where the shift hasn't gone as far as "PMs write the specs and ship the code."
What it means if you're a developer watching this happen
If review and verification are becoming a bigger share of what the job actually is, that's the exact skill worth deliberately protecting rather than letting atrophy — because it's turning into the differentiator, not a secondary skill next to "writes code fast."
Check how your own verification habits measure up →
Note: this trend is documented anecdotally across developer communities rather than through a single systematic study — treat the pattern as directionally real, not precisely measured. Skill-gap data cited from the Anthropic RCT referenced in Is AI Making Developers Worse?.
Related reading: Junior Developers and AI Dependency · Is AI Making Developers Worse?