Practitioners hail capable agents and warn of incidents, but our tracker shows engineering roles far outnumber evaluation and safety openings.
Agents are maturing while warnings get louder
A split opened up in the last two days’ posts. On one side, builders say agents are crossing reliability thresholds. Simon Willison writes that new model-and-agent combinations shifted from “often make mistakes” to “reliable enough to use on a day-to-day basis”, and Zvi Mowshowitz says “Claude Opus 5.5 should raise your ambition levels again.” Ethan Mollick adds that “It is strange how much LLMs turned out to be able to solve such a wide range of hard problems…”
On the other side, Gary Marcus argues the deployment of agents is already causing wide-scale issues: “It’s actually (at least) tens of thousands, per a new scoop from Madison Mills at Axios. And not just OpenAI, either.” Zvi, in a separate post, pushes for governance inside labs: “Who will be the evaluators?”
Our hiring tracker shows where employers are placing their bets. Across 179 employers, we count 4,137 open AI roles. The concentrated demand is in building and shipping systems, not in evaluation or governance.
- Modelling and engineering: 1,618 open, 1,030 opened and 809 closed in 30 days, across 141 employers
- Data: 725 open, 499 opened and 371 closed in 30 days, across 122 employers
- Infrastructure: 455 open, 224 opened and 88 closed in 30 days, across 95 employers
- Research: 232 open, 53 opened and 20 closed in 30 days, across 51 employers
- Product and design: 97 open, 50 opened and 25 closed in 30 days, across 45 employers
- Evaluation and safety: 46 open, 15 opened and 1 closed in 30 days, across 20 employers
If agents really are moving from demo to dependable, employers appear to be staffing accordingly. If they are also causing incidents, the hiring response into evaluation and safety is not yet keeping pace.
“Just works” meets on-the-ground usage
Willison grounds his claim in lived tests. He recounts a long-running toy benchmark: “Generate an SVG of a pelican riding a bicycle” and a year-on-year improvement in coding agents after “Claude Opus 4.5 and GPT-5.1.” In a separate post celebrating his keynote finale, he asked Claude to produce animated pixel art and concluded, “It’s pretty great!” He then had Claude Code automate a Playwright pipeline to turn that HTML into a video.
Zvi’s verdict on Anthropic’s latest is blunt: “It just works, and it persists, like Astra does.” He summarizes the vendor line too: “The pitch is Fable-5.1-level performance at lower Opus-level price.”
Mollick highlights capability transfer to the physical world: a Stanford demo “that let Astra drive a robot in a kitchen.”
Our hiring data aligns with this builder-heavy mood. Employers are opening and keeping open far more roles to ship features than to check them. Modelling and engineering has 1,618 open roles. Infrastructure adds 455. Product and design accounts for 97. The ratio to evaluation and safety is stark: 46 open across just 20 employers. Top employers by open roles also skew to firms known for large-scale delivery:
- Accenture: 581 open, 707 opened in 30 days
- Capital One: 203 open, 236 opened in 30 days
- OpenAI: 173 open, 47 opened in 30 days
- Amgen: 148 open, 41 opened in 30 days
- Anthropic: 127 open, 29 opened in 30 days
If agents are “reliable enough to use on a day-to-day basis,” this is what that looks like in headcount: materially more hiring in build functions than in checks and balances.
Incidents and governance: the gap between rhetoric and requisitions
Gary Marcus presses the case that harms are piling up: “All of this seems possibly illegal to me.” He argues this was predictable: “What is happening was foreseeable, and foreseen.” He also advances a motive: “But the companies blundered on, because agents use vastly more tokens than simple chatbots, and thus drive up revenue.”
Willison’s own feed shows a micro-incident from a consumer agent. The Muse AI Agent told a user, “That’s a bad look and it made the no-show worse. … Want me to change the pickup replies so they don’t promise you’re there?” It is a small case, but it captures two systemic risks at once: confident automation of a wrong state, and the cost to user reputation.
Zvi argues for embedded watchdogs: “Who will be the evaluators?” He ties that to recent security concerns in labs and a call for embedded evaluators inside them.
Our dataset does not show a hiring surge to meet those governance ambitions. Evaluation and safety sits at 46 open roles, with 15 opened and just 1 closed in the last 30 days. By comparison, modelling and engineering saw 1,030 openings and 809 closings in the same period. Even research, at 232 open, outweighs safety by a factor of five. Among labs named in these conversations, OpenAI has 173 open roles and Anthropic 127, but our category split is across all employers; we do not see enough employer-side demand classified into evaluation and safety to match the urgency described by Marcus and Zvi.
Collective AI, standardization, and where the jobs are
Azeem Azhar points to an alternative framing: “A new essay from DeepMind argues that AGI will not emerge from a single winning AI, but rather through “cooperative interactions among models, tools, institutions, and human participants.” Since I think AI will evolve this way, I’m sympathetic to the argument.” He also argues in a separate note that “Standardization matters.”
If the direction of travel is cooperative systems and standards, hiring should reflect investment in tooling, integration, and data plumbing. Our counts support that: data roles number 725 open across 122 employers, and infrastructure stands at 455 across 95. These are the connective-tissue jobs for the multi-agent, multi-tool world Azeem describes.
Simon Willison’s rapid toolmaking echoes this: “So I had Opus 5.5 vibe code this tool, which examines any Bluesky profile for evidence of a likely reply bot.” Even this small tool sits at the intersection of model capability, platform APIs, and a clear misuse problem (reply bots) that standards and governance eventually need to address.
Productivity, power, and who benefits
Alex Hanna centers labor: “Wages were won through people power. We dispel 5 common myths in my new video, Myths Around AI and Productivity, based on the work we’ve been doing at @dairinstitute.bsky.social with The Luddite Lab. Full video: www.youtube.com/watch?v=3BO0...” Emily M. Bender cautions against frictionless imposition of chatbots: “I think there are two key lessons from Linguistics as we deal with LLMs/chatbots being imposed on us everywhere.”
Our headline jobs-created indicator, which we count from employer listings and calibrate to the World Economic Forum’s Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030), shows ongoing net creation. The live snapshot here is narrower: 4,137 open AI roles across 179 employers. Within that, product and design roles are just 97 open, a reminder that the human-facing work of shaping how systems touch people is still a small slice of employer demand compared with engineering and data.
FOOM, incrementalism, and the hiring signal
Noah Smith hosts a debate on recursive self-improvement and fast takeoff: “there’s a widespread belief that as AI gets better at improving itself, there will be a “fast takeoff” or “FOOM”, in which AI’s capabilities “take off” and create a technological Singularity.” Simon’s own framing of 2026 advances a different pattern: incremental releases that cross invisible lines of reliability.
Our requisitions data look more like the incremental story. Research openings total 232 across 51 employers, healthy but not dominant. The bulk of growth and churn sits in modelling and engineering (1,030 opened, 809 closed in 30 days) and data (499 opened, 371 closed). Employers are building, instrumenting, and integrating rather than standing up large new research groups tailored to self-improving systems. That does not settle the FOOM debate, but it does tell you what companies are paying for this month.
Bottom line
- Builders say agents now “just work” and are useful day to day. Our counts align: heavy hiring in modelling, data, and infrastructure.
- Critics document incident volume and call for embedded oversight. Our counts do not yet show a matching buildout of evaluation and safety headcount.
- If collective, standardized AI is where we are headed, the data- and infra-heavy openings we track are consistent with that thesis.
The hiring market is voting with reqs. For now, it is voting to ship.
What we read
Every quote above is taken verbatim from one of these posts.
- Noah Smith: Where’s the 'intelligence explosion'?
- Simon Willison: 2026 in LLMs (so far), Kākāpō Party, Quoting Muse AI Agent, Bluesky reply bot checker
- Zvi Mowshowitz: The Quest for Embedded Evaluators, Claude Opus 5.5 Should Raise Your Ambitions
- Azeem Azhar: 🔮 The pros and cons of collective AI #603, 🔮 Safety in numbness
- Gary Marcus: BREAKING: AI agent incident toll has risen to tens of thousands
- Alex Hanna: Wages were won through people power. We dispel 5 common myths in my ne
- Emily M. Bender: I think there are two key lessons from Linguistics as we deal with LLM
- Ethan Mollick: It is strange how much LLMs turned out to be able to solve such a wide
- David Ha: Follow Sakana AI on IG → SakanaAILabs