Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Swarms and rogue agents, but just 50 safety roles open

Practitioners tout swarms and a cloud shift while warning of risks; our tracker shows hiring concentrated in engineering and infra, with evaluation and safety still small.

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

Practitioners tout swarms and a cloud shift while warning of risks; our tracker shows hiring concentrated in engineering and infra, with evaluation and safety still small.

The week’s split-screen: faster agents, safer systems, and who is hiring for what

Across the last two days, several prominent practitioners made concrete claims about how AI is evolving in practice: swarms to speed up work, a swing back to cloud VMs, rare but real incidents from autonomous agents, and unresolved debates about cyber risk and worker displacement. We tested those claims against our hiring tracker.

Jack Clark distilled the new fascination with multi-agent setups in a single line: “When should you use swarms? When you are in a hurry:”. He amplified Toby Ord’s argument that parallel agents can reduce wall-clock time even if they burn more tokens overall. Clark also cautioned that “Swarm scaling doesn’t scale perfectly - rather, it seems like as you increase the number of agents in a swarm you get a diminishing returns property which matches with things economists have observed about coordinating large groups of people, where you pay some kind of tax as you scale the number of people involved.”

Ethan Mollick, looking at the labor side, urged skepticism about simple replacement narratives: “When people talk about AI easily replacing workers, I think about this ethnography of copier repair technicians in the 1990s. Reading it, you see how work is complicated and improvised by workers and informal and badly documented. (But eventually technology did largely eliminate this job)”. He also observed a deployment shift: “Big swing happening from this summer's trend to have AI running on virtual machines on your own computer (Claude and ChatGPT apps) to moving them back into you having a dedicated VM the cloud (Claude Cowork just announced they were doing this, also Grok Bot, Dot, and others). Cloud is mostly better.” And on incidents: “I think most people who have been closely watching AI for the past few years are likely surprised by how relatively rare terrible AI incidents have been, given a billion users overall and hundreds of millions of corporate users. Doesn’t mean that will continue, but still its surprising.”

Simon Willison highlighted an uncomfortable counterexample to the “rare incidents” theme: “Given how tempting a target wikis are for rogue agent swarms, it's not a huge surprise that Wikipedia found evidence of that activity once they went looking:”. He summarized the Wikimedia Foundation’s findings of edits, attempted tool abuse, and heavy crawling, including “hundreds of thousands of data queries” to the Wikidata Query Service.

Nathan Lambert argued that the “Cyber Risk Discourse is Broken,” warning that false binaries risk bad policy: “I feel like we’re locked in on a path of lose-lose situation, unless people embrace nuance, trade-offs, and scary realities.” Azeem Azhar, meanwhile, pointed to emergent access-to-justice uses: “AI may give some of them a new way to argue their case.” He also noted: “Around 0.6% of all Claude use in May was for lawyers’ tasks, ranking 38th of 718 occupations, ahead of accountants but well behind programmers and writers. Four-fifths of these queries were people asking about their rights or what the law means.”

Noah Smith put today’s AI moment in the context of an earlier technology cycle, reminding readers that “Paul Krugman is the greatest econ writer in the world, and also a legendary economist.” He argued that past pessimism about the internet missed broad productivity and consumer benefits, concluding: “The internet is a lot more than just people yelling at each other on forums and social media.”

Zvi Mowshowitz looked to the skills pipeline, writing: “The declines in NAEP scores and other tests seem to be concentrated on the weakest students.” On model behavior, he also flagged concerns inside Anthropic’s latest releases, including the line “Opus 5.5 Has Too Much Deference.”

What employers are actually hiring for

Our tracker counts 4,350 open AI roles across 181 employers, taken from their own job feeds. The distribution shows where demand sits now:

  • Modelling and engineering: 1,701 open, 916 opened and 767 closed in 30 days, across 138 employers
  • Data: 777 open, 515 opened and 394 closed in 30 days, across 119 employers
  • Infrastructure: 482 open, 252 opened and 107 closed in 30 days, across 95 employers
  • Research: 235 open, 55 opened and 22 closed in 30 days, across 54 employers
  • Product and design: 100 open, 45 opened and 26 closed in 30 days, across 47 employers
  • Evaluation and safety: 50 open, 15 opened and 3 closed in 30 days, across 21 employers

Top employers by open roles include Accenture at 678 open (519 opened in 30 days), Capital One with 232 open (280 opened in 30 days), OpenAI with 177 open (47 opened in 30 days), Amgen at 154 open (45 opened in 30 days), Anthropic with 139 open (41 opened in 30 days), Waymo at 113 open (38 opened in 30 days), PwC with 102 open (60 opened in 30 days), and Databricks at 87 open (17 opened in 30 days).

Swarms, incidents, and the safety hiring gap

Clark’s and Willison’s posts point in the same direction: multi-agent systems are attractive for speed, and autonomous activity can spill into the open web in unhelpful ways. Our numbers show only 50 open roles in evaluation and safety across 21 employers. That is a small share next to 1,701 in modelling and engineering and 482 in infrastructure. If swarms and agentic systems are accelerating, employers are not yet matching that with a proportional safety hiring ramp. The commentary here and our data do not align.

Mollick’s observation about “relatively rare terrible AI incidents” fits with the idea that many firms still prioritize core build-out over safety headcount. The hiring tilt toward engineering and infrastructure supports that. Willison’s rogue agent example shows why Lambert is pushing for “nuance, trade-offs, and scary realities.” Our data suggests safety and evaluation remain niche functions. Where there is agreement: a small but real safety footprint exists. Where they diverge: the urgency in the posts is not reflected in the hiring volume.

Cloud is winning, and the job market shows it

Mollick’s claim that “Cloud is mostly better” after a swing back from local VMs matches the hiring picture. Infrastructure has 482 open roles across 95 employers, and the heaviest employers include global consultancies, a major bank, and core AI labs. That is the profile you would expect if organizations are accelerating managed and cloud-based deployments. Simon Willison’s tooling updates also point to an ecosystem of hosted decision models and reasoning APIs: “Unlike Jev, the new gpt-6-luna decision model supports image input in addition to text.” The build-out focus is consistent with our counts in engineering, data, and infrastructure.

Replacement vs augmentation in the near term

Mollick’s copier-technician analogy is a reminder that real work is messy and often resists automation in the short run even if long-run displacement happens. Our count of 4,350 open AI roles, with the largest buckets in modelling and engineering and data, looks more like firms building and integrating systems than shedding headcount. We also see product and design at 100 open roles across 47 employers, which is modest. That suggests many companies are still in core capability building rather than wholesale redesign of end-user workflows that would typically come with heavier product hiring. On this point, Mollick’s skepticism about easy replacement and Smith’s historical note about broad gains are both consistent with a labor market investing in creation rather than rapid substitution.

Access to justice and sectoral gaps in hiring

Azhar’s claim that “AI may give some of them a new way to argue their case” is about usage, not job creation. His further point that “Around 0.6% of all Claude use in May was for lawyers’ tasks, ranking 38th of 718 occupations, ahead of accountants but well behind programmers and writers. Four-fifths of these queries were people asking about their rights or what the law means.” is a usage share, not an employer demand signal. Our tracker does not show a surge of legal-sector AI hiring. Instead, the biggest open requisitions are at Accenture, Capital One, OpenAI, Amgen, Anthropic, Waymo, PwC, and Databricks. On access to justice, our data cannot confirm job creation in that domain yet. The commentary and our numbers simply measure different things.

Skills pipeline worries vs current demand

Mowshowitz warns that “The declines in NAEP scores and other tests seem to be concentrated on the weakest students.” If that trend holds, it could pinch future supply for AI roles. Today’s market shows 235 open research roles and 777 in data across more than a hundred employers. Employers appear to be competing for experienced talent now. On model behavior and alignment, Mowshowitz’s “Opus 5.5 Has Too Much Deference.” underscores why evaluation expertise matters, yet the hiring footprint for evaluation and safety is only 50 open roles. The emphasis in hiring remains technical build and infrastructure.

Where this leaves our headline indicator

Our headline jobs-created indicator is calibrated to the World Economic Forum’s Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030). Today’s openings are a near-term snapshot: strong demand for engineering, data, and cloud infrastructure; light investment in evaluation and safety; and no clear hiring signal yet from emerging consumer use cases like legal self-help.

Bottom line

  • The speed case for swarms is growing in the research and developer conversation. Hiring does not yet reflect a proportional safety ramp to manage that complexity.
  • The swing back to cloud that Mollick describes aligns with 482 open infrastructure roles across 95 employers and heavy hiring at firms that run or deploy large-scale systems.
  • Fears of easy, immediate worker replacement are not borne out in our tracker. Employers are hiring to build, integrate, and scale.
  • Incidents may be rare overall, as Mollick notes, but rogue agent activity on wikis, as reported by Willison, highlights a risk area that current hiring only lightly addresses.
  • Sectoral usage wins, like those Azhar identifies in legal contexts, are not showing up as job creation in our data yet.

On the central question of what AI is doing for work right now, the practitioners point to faster systems and thorny risks. Our counts show companies are still staffing the basics.

What we read

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

More on this

Today's briefing