--- title: "The job market looks stuck. AI engineering is where it isn't." excerpt: "The generalist software lane is crowded and slow. From a hiring seat, demand for engineers who operate models in production is loud and supply is thin." date: '2026-06-25' lastModified: '2026-07-04' author: 'Teo Deleanu' authorAvatar: '/team/teo.jpg' tags: ['AI Engineering', 'Career', 'Production AI', 'Hiring'] keywords: ['tech job market stuck', 'career change to AI engineer', 'AI engineer demand', 'production AI skills'] featured: false tldr: 'The generalist software lane feels stuck for structural reasons: coding agents made typing cheap, so a crowded queue now competes on exactly the input that lost its price. From a hiring seat the production-AI lane looks opposite: every AIESCU engagement is founder-direct, the asks are production skills, and our own 90-day plan has to hedge its two-senior-offers-by-day-45 commitment because that profile is hard to source. The bridge between the two markets is finite: the six failure surfaces of production AI, proven by one narrow system you build and operate. This is a lane change for people who can already code, and it asks for an operated system rather than a certificate.' keyTakeaways: - 'The stuck feeling is a lane problem. The generalist queue is crowded while the production-AI operator lane is thin on supply, and the same engineer can stand in either.' - 'From our hiring seat, sourcing senior production-AI engineers is hard enough that our 90-day plan hedges its own funnel: a warm network, three parallel sourcing channels, a rubric written in-house.' - 'Coding agents made typing cheap. The market now prices the judgment layer: review, architecture calls, operations, and accountability for incidents. That layer is the AI-engineering skill set.' - 'The bridge is finite: six failure surfaces, learnable in months, proven by one narrow system operated in production rather than by certificates.' --- One profession, two markets. In the lane where most software engineers are searching, the story I keep hearing is hundreds of applications and long stretches of silence. In the lane where I hire, the story is a 90-day plan that commits to two senior production-AI offers by day 45 and a risk register that flags the hiring funnel as the item most likely to slip. Same profession, opposite problem. I sit on the demand side of the second lane. AIESCU's engagements are founder-direct, and part of what I sell is hiring itself: writing the rubric, sourcing candidates, running interviews, and setting the technical bar. This post is what the market looks like from that seat, and why I think the bridge between the two lanes is a finite, learnable skill set. ![Two market lanes for the same engineer: a crowded generalist lane with long queues and automated screens, a thin-supply production-AI operator lane with founder-direct demand, and a bridge listing the six failure surfaces as the finite skill set between them](/blog/two-markets-one-gap.svg) ## Why the generalist lane feels stuck This section stays qualitative on purpose. Most numbers passed around about the tech job market do not survive a source check, so I will only report the pattern engineers describe to me: hundreds of applications for a mid-level generalist role, screening software deciding before a human reads anything, interview loops that end in a form letter, recruiters who vanish after the first call. If that matches your last six months, the problem is the lane, and the lane is crowded. Part of the crowding is structural, and I have a first-hand receipt. The [live demo](/demo) on this site went from written spec to deployed system in 29 days and 74 commits, one engineer plus coding agents, a build documented in [the consultancy post](/blog/one-person-consultancy-four-weeks). Read that as a labor-market fact rather than a brag: typing code stopped being the scarce input. A team that once needed several generalist implementers can now ship with fewer, and a resume whose core claim is "I implement features from tickets" competes directly with the tooling that compressed my build to 29 days. What stayed scarce is the judgment layer: reviewing the change, making the architecture call, operating the system, and answering for the incident. On my own build, review bandwidth was the only constraint left. That layer is what the AI-engineering role packages, which is why the same engineer can feel unwanted in one lane and get recruited hard in the other. ## What demand looks like from a hiring seat Every AIESCU engagement is founder-direct, so I hear the demand without a filter. When founders describe the engineer they need, the list is production-shaped: retrieval that finds the right context, evals wired into CI, observability that catches failures returning HTTP 200, a deploy that survives the provider's next deprecation. Model theory almost never comes up in those conversations. Founders are asking for people who can operate the system around the model. The hiring side is the sharper receipt. The 90-day plan in [the operator brief](/about) commits to two senior production-AI offers out by day 45, sourced through the AIESCU network and production-AI contacts, with a written and calibrated hiring rubric behind both offers. The same plan lists its named risks, and the hiring funnel is on that list: the 45-day timeline assumes a warm network, so three sourcing channels run in parallel from day one. Read the hedging as a market signal. When the person selling the hiring plan protects his own funnel this carefully, supply is thin. Sourcing senior production-AI engineers is a deliverable I am paid for, and it is hard. Public signals point the same direction. LinkedIn's 2026 Jobs on the Rise report ranks AI engineer the fastest-growing job in the US, with four of the top five roles tied to AI. I trust my own funnel more than any ranking, but the two agree. ## The gap exists because the skill set is new and production-shaped Nobody planned this gap. The role is too new for the training pipeline: CS programs teach algorithms and bootcamps teach frameworks, while the job defends a production system at the places where models meet reality. Those places are countable. On the systems I operate, six surfaces cover most of what breaks: retrieval, agent loops, evals, observability, deployment and operations, and coordination at scale. Each one earned its spot through a documented incident, and the full curriculum argument, with a receipt per surface, is in [becoming an AI engineer](/blog/becoming-an-ai-engineer). This post is the why-now. That one is the how. Finite matters more than easy here. The list is short enough to learn in months, and it has held still: the model layer changed quarterly across every migration we ran, while the six surfaces stayed put. A skill set that survives model churn is a reasonable thing to point a career change at. A prompt trick carries no such warranty. ## What filling the need concretely means One narrow system, built end to end and operated in production, beats any stack of certificates. Certificates say you watched. An operated system says you were on the hook. [The narrow-loop playbook](/blog/narrow-loop-playbook) is the sequence I recommend: pick one loop with a measurable output, write the golden questions before the code, ship the thinnest complete pipeline to real infrastructure, gate merges on your eval delta, then run it for a week and read the logs daily. The demo on this site is that path executed, and the operating week mattered more than the build. Its first seven days in production included a model retired upstream, with 404s as the notice period. Surviving that week produced numbers I can defend in any interview: what recall did when the chunking changed, what a query costs, what broke in week one, and what the fix was. A hiring rubric like mine screens for exactly those numbers, because they cannot be produced by watching tutorials. ## Who this is not for An honesty section, because the how-to sibling of this post makes promises I intend to keep. If you cannot already write and debug programs, this lane change is premature. AI engineering is systems engineering around models, and the six surfaces assume you can build the system they defend. Learn programming first. The surfaces will still be there. If you want the title without operating anything, this market will find you out. Founder-direct hiring means the founder watches what you do in week one, and reference checks in a thin market are short paths between people who know each other. The rubric I write asks what you ran, what broke, what you changed, and what the numbers did afterward. A retitled resume answers none of that. And if you want certainty, I have none to sell. Urgency is a sales tactic, and I will not run one here. What this market rewards is an operated system, whenever you show up with one. ## The distance is six surfaces wide The stuck feeling in the generalist lane is real, and I will not talk you out of it. The claim of this post is narrower and checkable: from where I sit, writing the rubrics and working the funnel, demand for production-AI operators is loud, supply is thin, and the distance between the two lanes is six learnable surfaces plus one system you run yourself. The standard the demand side holds is written down in [the operator brief](/about), the system that standard produced is answering queries at the [live demo](/demo), and if you want the guided version of the bridge, [the DIY program](/diy) teaches the six surfaces on a system you build yourself. Start the narrow system. The queue will still be there if you change your mind.