Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Agents get the hype, but safety hiring stays tiny in our data

Practitioners tout agents, open models and breakthroughs; our tracker shows hiring still dominated by engineering and integrators, not safety or product.

Read the original at AI Jobs Report
9 min read27 viewsBy AI Jobs Report

Practitioners tout agents, open models and breakthroughs; our tracker shows hiring still dominated by engineering and integrators, not safety or product.

A week of agents, governance, and a reality check from jobs

Across the last two days, several prominent voices sketched a world where AI agents get cheap and ubiquitous, governance choices stay undecided, and deployment accelerates. Azeem Azhar says, "Companies are backing their AI claims with more numbers." Simon Willison highlights TypeSafe’s Jev, a model class he frames as decision-oriented, writing "It's also very fast, and really cheap." Zvi Mowshowitz asks, "What should your AI lawyer do for you?" while Gary Marcus warns that "Hope is not a strategy."

Our tracker counts 4,319 open AI roles across 178 employers. The pattern in the data is clear. Hiring remains concentrated in modeling and engineering and the integration layer, not in safety or product. That puts boundaries around some of the expectations on show this week about agents and governance.

Agents get cheaper and faster, but who is hiring to productize them?

Willison describes Jev as a shift in interface and cost profile: "Jev charges only for input - output is free - and the input price of their first model is $0.042 per million tokens - cheaper even than OpenAI's GPT-5 Nano ($0.05/million)." He calls it a new category and quotes pricing that, if widely matched, could push deployment decisions toward many small, embedded agentic judgments rather than long chats.

Ethan Mollick points to accessible agent experiences already in mainstream products: "I think Meta's Muse is an impressive implementation of the OpenClaw idea of AI as a personal assistant agent that you have an ongoing chat with." He also shows non-coding use cases landing for enthusiasts: "AI can be a really wonderful tool for exploring topics far from coding." and later, "As an Eliot fan, I am quite impressed."

Our hiring data does not show a corresponding surge in productization roles that you might expect if agent UX and decision model workflows were the main bottleneck. Product and design roles are 109 open, with 54 opened and 14 closed in 30 days, across 49 employers. By contrast, modeling and engineering dominates at 1,674 open, with 1,021 opened and 607 closed in 30 days, across 142 employers. Infrastructure also outweighs product, with 444 open, 199 opened and 68 closed in 30 days, across 97 employers.

This pattern agrees with Willison’s broader developer lens that platforms are moving to make AI easier to embed. He notes Cloudflare’s expansion of the execution surface for Python: "Python is now a first-class, fully supported language on the Cloudflare Developer Platform" and details how this lands via Pyodide and Workers. Our numbers line up with that kind of buildout. Infrastructure demand is meaningful and widespread across employers, while product headcount remains comparatively small.

Governance talk is hot. Safety jobs are not.

Jack Clark summarizes RAND’s posture on uncertain takeoff: "RAND thinks the best AI strategy for the US is to keep all options open:" and quotes the thrust as "It’s not clear exactly what is optimal, so the US should spend money now to preserve options…" Marcus, addressing the UN, rejects both laxity and fatalism, writing "We are at a moment of moral, political, and economic crisis" and cautioning against extremes. His near term risk callout is specific: "What we are facing near-term, anyway, is not extinction. And it’s not superintelligence… It’s wholesale deepfaked disinformation, and unreliable but persistent AI systems stealing credentials and launching cyberattacks, at scale."

Mowshowitz narrows to the question of loyalty and limits in legal assistants: "If I’m being sufficiently evil then I sure hope it tells me no." He explicitly brackets superintelligence to talk product design in the tool era: "This post is about a non-ASI ‘AI as mere tool and normal technology’ world."

Our tracker shows that despite the intensity of governance and misuse concerns in commentary, hiring for these safeguards is the smallest category we track. Evaluation and safety roles total 48 open, with 17 opened and 3 closed in 30 days, across 21 employers. That is a fraction of modeling, data, and infrastructure hiring.

Where the commentary and the data agree: engineering remains the center of gravity. Where they diverge: the public conversation emphasizes safety, evaluation, and policy tradeoffs, but employers are barely staffing those areas compared to buildout roles. If new agent modalities and governance challenges were already driving a rebalancing toward assurance, our data would show more than 48 open evaluation and safety roles.

Open-weight positioning, but hiring is about build capacity

Nathan Lambert sets the competitive frame he briefed to Congress: "I was recently invited to brief a group of Congressional members and staff on the state of open-weight models in the lens of U.S.-China competition." He distinguishes open-weight from open-source and closed access, offering a clean definition: "Open language models are AI models where their weights are publicly available for inspection or downstream use." That lens highlights license and ecosystem dynamics, and it is a debate that often spills into workforce questions.

Our dataset does not tag roles by open-weight versus closed model stacks, so we cannot test that split directly. What we can see is who is hiring and for what. The top hiring employers are an integrator and major enterprises rather than a single open-weight champion. Accenture tops the list with 649 open, 748 opened in 30 days. Capital One has 192 open, 179 opened in 30 days. OpenAI lists 175 open, 48 opened in 30 days, and Anthropic 128 open, 31 opened in 30 days. That mix points to a market still dominated by implementing and integrating models, whatever the licensing, rather than a decisive tilt toward any one open-weight play.

Platforms spread, workflows lag

Willison also reported that Cloudflare’s Python Workers are now GA, and he credits maintainers pushing Pyodide forward. That aligns with our observation that infrastructure listings are broad based across 97 employers. But the practitioner accounts hint at a process gap. Willison quotes a practitioner, voxium, embedded in a large company: "Nobody on my team likes this. They are being forced to ship as much as they can." and the starker line, "People are working 12 to 13 hours a day just to press enter."

Our breakdown suggests organizations are still prioritizing hiring for builders over people tasked with workflow and outcome quality. Data roles are 778 open, with 511 opened and 298 closed in 30 days, across 129 employers. That is a large, necessary foundation. But the small product and design count noted above, and the equally small evaluation and safety count, are consistent with voxium’s complaint that process and review can be an afterthought. The posts tell a story of rapid deployment. The jobs show companies still staffing the pipes, less so the brakes and steering.

On the data governance side, Alex Hanna highlights the academic focus on data practices: "Today @emilymbender.bsky.social noted that our paper on data practices in machine learning with @rajiinio.bsky.social, Amandalynne Paullada, and Remi Denton just passed 1,000 citations." Our data category is large, but it bundles many functions and we cannot isolate governance within it. So while the research discourse is maturing, we cannot yet show a distinct hiring spike for data governance from our role taxonomy.

Breakthrough claims land, but integrators hold the most roles

Mollick relays a striking claim about frontier progress: "OpenAI "has now resolved more than 100 long-standing open problems across most areas of mathematics," and is waiting to release them until after discussions with the math community" and speculates about spillovers to professional certification. Our tracker cannot validate research impact, but we can quantify who is hiring. Even with labs like OpenAI and Anthropic appearing in our top employers, the largest volume is at Accenture and other enterprises where the work is less about beating math benchmarks and more about rolling out systems at scale.

That matches Azeem Azhar’s focus on earnings call signals and the wider economy. It also fits Jack Clark’s summary that strategy is to keep options open. Enterprises appear to be doing just that, staffing engineering and data-heavy roles that keep options open for multiple model providers and deployment patterns.

Where commentary and hiring meet next

A few tests to watch in our future snapshots:

  • If governance concerns are turning into headcount, our Evaluation and safety category should grow from 48 open across 21 employers.
  • If decision models and agents are bottlenecked by UX and policy, Product and design should climb from 109 open across 49 employers.
  • If platformization continues, Infrastructure could rise from 444 open across 97 employers alongside integrators like Accenture, which already lists 649 open.

Our tracker counts job listings from employer feeds. Our headline jobs-created indicator is calibrated to the World Economic Forum’s Future of Jobs Report 2025, which projects 11M AI roles created and 9M displaced by 2030. Against that long horizon, this week’s posts capture the tension between big claims and near term execution. The hiring market, as measured by us today, is voting for build capacity first. Safety, evaluation, and productization remain the smallest slices, despite how much they dominate the conversation.

What we read

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

More on this

Today's briefing