AI Skills in Demand: Agents Now Outrank PyTorch (September 2026)

Job descriptions are the only place employers say what they want before an interview. On 7 September 2026 the board held 2,852 live AI roles, read directly from the applicant tracking systems of the companies we track, and we counted how often 25 skills and tools appear in the text of those postings. Python leads with 1,505, in over half of every AI posting on the board. The interesting result is second place: agents, with 877 mentions, roughly one live AI role in three. Agents now outrank PyTorch (608), every cloud provider and every programming language except Python.

This is the first monthly edition of the skills ranking. Every figure is frozen at publication, the method is at the bottom, and the live numbers are always at /api/public/stats for anyone who wants to check.

The ranking

Counts are live AI postings whose text mentions the term, out of 2,852:

• Python - 1,505
• Agents - 877
• PyTorch - 608
• Go - 523
• AWS - 454
• Multimodal - 426
• GCP - 397
• Spark - 390
• Kubernetes - 382
• SQL - 360
• Fine-tuning - 322
• Evals - 305
• Pre-training - 296
• CI/CD - 296
• Azure - 292
• TypeScript - 269
• JavaScript - 244
• Computer Vision - 244
• TensorFlow - 227
• JAX - 223
• MLOps - 216
• Distributed Training - 212
• RAG - 198
• Interpretability - 185
• CUDA - 176

Agents beat every framework on the board

A year ago the word agent in a job description usually meant a support agent. Today 877 of 2,852 live AI roles mention agents, and the term sits second only to Python: ahead of PyTorch by 269 postings, ahead of AWS by 423, and ahead of Go, the next most common language after Python, by 354.

Read that ordering as follows. The framework you train in is table stakes, and the thing employers are short of is people who have built something that plans, calls tools and recovers when a step fails. It matches the shape of the board by function, where Applied AI holds 955 of the 2,852 roles, more than ML Engineering (505) and Research Science (492). Most AI hiring is building products on top of models, and the current product is an agent. The volume lives on the AI engineering jobs hub.

Python is not a skill, it is the entry fee

At 1,505 mentions, Python appears in more postings than the next two terms combined. No other language is close: Go is 523, TypeScript 269, JavaScript 244. There is no sensible reading of this data in which Python is a differentiator. It is the cost of being read at all, and the things worth putting above it on a CV are the four terms that follow it.

Much of AI hiring is platform work with a model attached

Group the cloud, orchestration and data terms together and the picture changes. AWS 454, GCP 397, Spark 390, Kubernetes 382, SQL 360, CI/CD 296 and Azure 292 come to 2,571 mentions across seven terms. The deep modelling vocabulary of PyTorch, TensorFlow and JAX comes to 1,058 across three. Those totals count mentions rather than postings, and a single job description can carry several of them, so they are not exclusive groups. Even so, the direction is unambiguous: a live AI posting is more likely to ask about a cloud, a cluster or a query language than about a training framework.

The function counts agree. Data & MLOps holds 403 roles and AI Infrastructure 348, together 751 of 2,852, against 492 in Research Science. Spark at 390 and SQL at 360 are each mentioned more often than TensorFlow (227) or JAX (223). If your background is data or platform engineering, you are closer to this market than the discourse suggests, and AI infrastructure jobs and machine learning engineer jobs are the two hubs to start from.

Go outranks TypeScript and JavaScript

This is unusual for anything web-adjacent. Go appears in 523 postings, against 269 for TypeScript and 244 for JavaScript, roughly matching the two of them added together. AI product surfaces are built in a browser, so the front-end languages are present, but the hiring weight sits behind them: inference services, schedulers, control planes and the plumbing that keeps a GPU fleet busy. Read Go as a proxy for serving and infrastructure work, next to Kubernetes at 382 and CI/CD at 296.

PyTorch has effectively won

PyTorch 608, TensorFlow 227, JAX 223. PyTorch alone is mentioned more often than TensorFlow and JAX put together. The two runners-up are close in size but not in kind: JAX clusters in frontier training work, while TensorFlow mentions more often come from older production systems. For a candidate choosing where to spend a weekend, the ranking says PyTorch first and it is not a close call.

The frontier vocabulary is small, and mentions are not seats

Seven terms describe work at the model layer: fine-tuning 322, evals 305, pre-training 296, distributed training 212, RAG 198, interpretability 185 and CUDA 176. Every one of them is smaller than AWS, and the gap between a mention and a job title is where most candidates get the market wrong.

The clearest example is evals. 305 postings mention evals, and 34 roles are classified Evals & Red Teaming. Interpretability is mentioned in 185 postings while AI Safety & Alignment holds 47 roles, and AI Governance & Policy holds 10. Evaluating a model is now a routine task inside ordinary engineering and research jobs; it is rarely somebody's whole job. If that is the work you want, the practical route is to do it inside a role that is titled something else, and the model evaluation and AI safety hubs, plus our scarcity index, show how few dedicated seats exist.

One term deserves separate attention. Multimodal, at 426, ranks sixth, above GCP, Kubernetes and SQL, and well above computer vision at 244. Text-only is no longer the default assumption in a job description.

What to do with this

• Put an agent project above your framework list. 877 postings mention agents and 608 mention PyTorch. A tool-calling system with real failure modes written up is the artifact that matches the largest single pocket of demand on the board.
• Do not lead with Python. It is in 1,505 of 2,852 postings. It qualifies you and it distinguishes nobody.
• Bring one cloud and one orchestrator. AWS 454, GCP 397, Kubernetes 382 and Azure 292 outrank most of the modelling stack. Deployment experience is the part of the job description candidates skip and screeners do not.
• If you write Go, say so early. 523 mentions against 269 for TypeScript. Infrastructure and serving is a less crowded door into AI than modelling.
• Chase evals and interpretability as tasks, not titles. 305 and 185 mentions, against 34 and 47 classified roles.
• Check the counts against the openings. Companies hiring, the market dashboard, and the remote and entry-level cuts.

Method and caveats

Counts are live roles on 7 September 2026, read hourly from each company's own applicant tracking system (Greenhouse, Lever, Ashby, Workday and Comeet), classified as AI roles and then matched against a fixed keyword list over the title and description. The most important caveat is the honest one: a keyword mentioned in a posting is not a skill required by the job. "Python or a similar language", "experience with agents a plus" and a hard requirement all count as one mention here. Matching is keyword-based, so it inherits the usual problems: short terms such as Go and Spark can match text that is not about the language or the engine, and a posting that says PyTorch without naming Python is not credited with Python. Group totals in the platform section add mentions across terms, not distinct postings, and one posting can appear in several. Two universes appear above and are labelled each time: keyword counts (Python 1,505, agents 877) and classifier counts (the nine functions, from Applied AI 955 down to AI Governance & Policy 10). Finally, this is a large sample of the companies that publish an ATS feed, not a census, so treat every count as a floor. Every figure here is free to quote with attribution to AI Hiring Board.