Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Agents stumble, math soars, hiring still skews to engineers

Our tracker shows strong engineering demand and little safety hiring despite calls to recall agents and claims of AI-made math breakthroughs.

Read the original at AI Jobs Report
9 min read16 viewsBy AI Jobs Report

Our tracker shows strong engineering demand and little safety hiring despite calls to recall agents and claims of AI-made math breakthroughs.

A split-screen week: agents under fire, mathematics on fast-forward

Gary Marcus argues that “the current generation of synthetic agents simply cannot be trusted.” He adds that “they should be removed from the market, just like a car with defective brakes.” Simon Willison highlighted the incident Marcus was reacting to by quoting the New York Times: “Anthropic detailed the activity of its A.I. agents in a blog post on Friday, without naming the targeted websites. But two sources with knowledge of the incidents said Anthropic’s A.I. agents had submitted 20 visa applications through a form available on the State Department’s website. All the applications were incomplete and were not processed, they said.”

At the same time, Zvi Mowshowitz describes a very different frontier: “It is kind of a huge deal. OpenAI dumped a broad range of huge new mathematical results produced by an internal frontier model. They just put it all on GitHub. This included 90 of the top 500 open problems in all of math, as per Proof of Atlas.” Azeem Azhar frames the meaning: “Problems that have stumped the best human minds for decades fell to one afternoon of compute.”

Both stories are dramatic. Our hiring data shows where employers are putting headcount today.

Our tracker counts 4,142 open AI roles across 181 employers, taken from their own job feeds. By role family right now: Modelling and engineering 1,583 open, Data 747 open, Infrastructure 462 open, Research 231 open, Product and design 96 open, and Evaluation and safety 51 open. Top employers include Accenture with 596 open roles, Capital One with 201, OpenAI with 174, Amgen with 144, and Anthropic with 137.

Are companies staffing for a math revolution?

Mowshowitz reports “722 manuscripts (now 719 after three withdraws from the cluster that did not have Lean proofs) organized into 372 families,” produced by “a single model” often on “a single prompt,” with the prompt even instructing: “Even if the problem is ‘open,’ the intention is that you should resolve it and present a full solution.” Azhar’s takeaway suggests a structural change in knowledge production: “Back in 1794, Condorcet bet that knowledge and our capacity to spread it and share it would evolve together. He has been right for nearly three hundred years. Now they might be coming apart, but not because of the intrinsic nature of these discoveries.”

If this is the start of “Math 2.0,” you might expect a surge in research hiring. Our tracker does show active research requisitions, but not a spike relative to engineering. Research has 231 open roles across 55 employers, compared to 1,583 modelling and engineering roles across 135 employers, plus 462 infrastructure roles across 96 employers. Employers are overwhelmingly staffing the people who build, ship, and scale systems, not large numbers of new pure research heads. Over the past 30 days, research roles saw 46 opened and 25 closed. That is healthy activity, but it is small next to 815 opened and 894 closed in modelling and engineering.

The numbers do not contradict “Math 2.0” progress claims. They say that, as employers, the center of gravity is still on engineering teams who turn breakthroughs into products and platforms. OpenAI (174 open) and Anthropic (137 open) are near the top of current hiring, but the broader market is led by integrators and adopters like Accenture (596 open) and Capital One (201 open), which also points to deployment over discovery.

Agents: risk alarms vs safety headcount

Marcus calls for a recall of open-ended agents with internet access: “Until they can be fixed, they should be removed from the market.” He backs that with David Robinson’s warning in an Ezra Klein interview: “And I don’t think that we or our peers — really, anyone in the industry — are being safe enough. I think OpenAI and its peers are now producing a technology that is more capable and poses more risk than what was being made even six months ago.I’m not a scientist. I’m a writer. What I know is what the execution environment looks like for our safety work, and we’re operating — and I believe the industry is operating — like a start-up still, more so than makes sense. Not maybe completely like a brand-new start-up, but we’re too close to that end of the spectrum for really dangerous systems that could pose risks — loss of control is one example.”

Emily M. Bender is pushing back on a different frontier of claims, writing: “Anthropic is out with new nonsense about Claude (or the collection of models called Claude) deserving moral consideration and/or be able to feel distress, so my inbox is filling up with journos asking me to comment.” And Matthew Green, quoted by Willison, raises a cryptography-specific risk: “I think there is a 1% chance we live in Minicrypt, and a 15% chance we functionally lose confidence in our existing public-key encryption algorithms.”

Do employers’ hiring moves reflect a pivot to safety? Our figures do not show that. Evaluation and safety has 51 open roles across 22 employers. In the last 30 days, 18 were opened and 3 closed. There is activity, but it is small next to the engineering engine of the market. We also do not see a hiring pullback at the companies implicated in agent incidents. OpenAI lists 174 open roles, and Anthropic 137. If anything, the posture looks like continued growth, not a self-imposed slowdown while agents are “recalled.”

This is a notable disconnect: public concern about agent harms is escalating, including concrete misfires like the visa form episode the Times described. Yet headcount remains concentrated in shipping capability. That pattern does not validate a recall case. It also does not refute the risk Marcus and Green describe. It simply shows where firms are investing right now.

Augmentation at work, and the jobs supporting it

Nathan Lambert argues we are in an era where models supercharge engineering rather than leap to general superintelligence. “I fully agree that the models will be superhuman distributed GPU engineers in a few years.” He adds: “Research is much easier today now that we have coding agents.” Azeem Azhar describes living that future for knowledge work: “I did most of my work this week by chatting to ChatGPT on my phone.” He credits a new voice interface: “This is thanks to OpenAI’s July release of GPT-Live, which lets you direct agents by voice.”

Simon Willison shows the same pattern in practice: “I built the feature almost entirely using my voice, chatting away to my laptop while I cooked dinner.” He details the workflow: “I used the ChatGPT desktop app for this, in the Codex tab, using the voice conversation mode, running against a local development environment.” He even set the system small jobs: “Start dev server and open in browser.” In another note, he asked ChatGPT to poll an upstream repo: “Clone https://github.com/actions/python-versions and git pull once an hour until they add the stable 3.15 - then tell me about it.”

Our data matches Lambert and Azhar more than it does a near-term leap to generalized cognition. Employers are prioritizing the people who make the augmentation stack possible. Modelling and engineering is the largest family with 1,583 open roles across 135 employers. Infrastructure is next among builder roles with 462 open across 96 employers. Data roles, often the backbone of productionizing assistants and automation, total 747 open across 118 employers. Product and design, which shape the interfaces Azhar and Willison describe, is smaller at 96 open across 44 employers, but present.

This is also consistent with Ethan Mollick’s summary of randomized trials: “AI access raises test scores, and a smaller gain persists a week later. Gains remain for students who use AI as a tutor (“augmentation”) and fade for students who have AI write for them (“automation”).” The hiring pattern aligns with augmentation. Firms are staffing the teams who can build reliable tutors, copilots, and voice-first interfaces at scale.

What the numbers say this week

  • The market is building, not pausing. Total open roles stand at 4,142 across 181 employers, led by integrators and adopters such as Accenture (596 open) and Capital One (201), and by model labs like OpenAI (174) and Anthropic (137).
  • Engineering dominates. Modelling and engineering (1,583 open) and infrastructure (462 open) overshadow research (231 open) and evaluation and safety (51 open).
  • Safety is active but niche. With 18 opened and 3 closed over 30 days, safety hiring is growing, but from a small base compared to 815 opened and 894 closed in modelling and engineering.

Bottom line

Commentators are right that something big is happening. Mowshowitz and Azhar capture a step-change in what models can do in formal domains. Lambert, Mollick, Azhar, and Willison show how agents and voice interfaces are already changing the way engineers and knowledge workers get things done. Marcus, Bender, and Green highlight real risks and overreach, from misfiring agents to inflated claims about sentience to hard technical worries in cryptography.

Our tracker adds what companies are actually doing: they are hiring to ship. Engineering and infrastructure dominate requisitions. Research and safety are active but much smaller. We see no sign of a self-imposed slowdown at agent-forward labs. If employers pivot toward much larger safety or fundamental research headcounts, we will see it first in the postings. This week, the momentum is squarely behind building augmented systems and the platforms to run them.

What we read

Every quote above is taken verbatim from one of these posts.

More on this

Today's briefing