One darkly amusing facet of the agentic revolution is how proudly and triumphantly thought leaders are sharing "revelations" that have been common knowledge for decades.
I was inspired to write this by a TianPan.co blog post on SWE laddering, but only because it most recently crossed my desk. It isn't uniquely egregious, but it's short and representative, and I can easily link it, so it's a good example. Its thesis is reasonable prima facie:
Teams that built their career ladders for deterministic software systems are applying those same frameworks to AI engineers — and systematically misidentifying their best people.
The article describes why "standard" (e.g. pre-agentic) leveling schemes won't work in a "probabilistic" world. Here's its presentation of "AI engineer" leveling, but with a twist: I've elided the AI-specific language. See what you notice.
Junior...engineers are essentially integrators. They call APIs, implement...pipelines, wire up existing components, and handle data preprocessing. They follow guidance on...design. They write the features someone else architected. The ceiling of independent judgment is intentionally low — not because they can't think, but because they haven't yet built the intuition for where...systems fail in ways that look like success.
Mid-level...engineers start designing systems, not just implementing them. They build complete...pipelines.... They make the first-order decisions.... Critically, they're comfortable operating with ambiguity — they can take a vague product requirement and convert it into a concrete technical plan without a blueprint. The benchmark for promotion isn't "can they write harder code" but "can they decompose ill-defined problems into defined ones."
Senior...engineers are doing something fundamentally different: they're establishing the ground truth. They define what "good behavior" looks like for the system. They design the evaluation framework that everyone else's work gets judged against. They make the architectural calls that constrain every feature built on top. When something goes wrong in production that no one anticipated, they're the ones who can trace backward from a failure mode to a root cause and forward to a systemic fix.
Staff and above operate at organizational scale. They're not just building...capabilities — they're building the infrastructure of judgment. Governance frameworks for...version[ing] and audit[ing]. Evaluation standards for when a[n]...upgrade ships. The criteria by which the organization decides what...can and cannot be trusted....
That looks great to me. But it looked great five, ten, twenty years ago. This is bog-standard SWE leveling. Agentic work doesn't change its fundamental nature.
Why does the author see a fundamental shift? They tell us by telling us what's wrong with pre-agentic leveling:
Velocity as a proxy for output
Code complexity as a proxy for skill
Benchmark scores as a proxy for production quality
Misreading the adoption learning curve [i.e. not letting seniors take time to ramp up]
These are certainly signs of bad leveling - but they're not newly bad. These flaws have been known for decades (although they are, regrettably, still endemic). The article has mistaken anodyne mismanagement for obsolete determinism: Its criticism of pre-AI SWE ladders misidentifies these ladders' flaws as incompatibilities with AI-native software development specifically rather than the general leveling weaknesses they actually are. Yes, they don't work in the new world. They didn't work in the old one, either. The problem is reprised in the article's proposed ladder for AI engineers, which is a good ladder for every SWE once you take the AI-specific stuff out. AI engineering is just engineering.
I'm seeing this broad confusion all over the industry. Accurate internal documentation, rigorous monitoring, multi-modal information availability, and well-defined standards of correctness have always been tremendously valuable - but only now that agents need them are tech executives starting to acknowledge this value. I recently heard of a CTO who proudly announced his recent executive decision that planning is the most worthwhile thing his engineers can do, and decreed that they should focus more on that than writing code. This sounds like a mean parody, but the materials I've seen make me confident that my account is faithful. Apparently this announcement came shortly after the release of Cursor's "plan mode" feature, which suggests the shocking possibility that this CTO thought that technical planning as a concept was a novel feature of agentic development rather than how competent engineers use their brains every day.
Optimists might be buoyed by this - after all, good decisions (e.g. fixing SWE performance ladders or dedicating time to rigorous quality assurance) for bad reasons are still good decisions. But I do not believe that these proclamations constitute "decisions" in any substantive sense. I think they're all just saying this stuff. An individual leader here or there might effect some particular change in her organization that she ascribes to agentic adoption, and that change might even be good, but I don't predict the industry broadly improving by any metric.
The main reason for my skepticism is the failure of agentic development to deliver any broad industry changes beyond massive psychic damage to line engineers. We are constantly reading executive announcements that the world has changed, each worded almost identically, each suspiciously unpaired with any visible consequence. Well-known companies won't shut up about their agentic transformations, but they're doing the same things they always were, with the same headcount, at the same delivery rate, for the same price. (I am not the only one who thinks this - ask a consultant who deals with these companies regularly, or, you know, the COO of one of them.) I think they're all reading from the same speaker notes because tech leadership are some of the most impressionable people in the world (due to their characteristic blindness to material reality), and AI mass hysteria is perfectly calibrated to attack the weakest parts of their cognitive defenses. But nothing is changing beyond getting more obnoxious.
I don't know where this is all going to go. The economics of it don't seem sustainable, but a bubble popping won't send us back in time. Something will end up different. Just don't trust some guy on LinkedIn to predict its particulars. If he knew them, he'd be doing more than posting.