AI Agents, Prices, Jobs: What Really Matters Today
New research on agents, cheaper models, Anthropic’s job fund, and a hacker attack with AI agents: the daily AI Radar digest with context.
Inhaltsverzeichnis
Today is one of those days when AI sounds at once like a research lab, a trading floor, and a security operations center. Between new agentic approaches, falling model prices, and real consequences for work and cybersecurity, it’s becoming clear: the market is not just changing the technology, but also the rules around it.
If you take away just one thing: AI is getting cheaper, more autonomous—and not automatically safer. Pretty inconvenient, but relevant.
🤖 World Action Critic: When agents evaluate their own actions
With World Action Critic (WAC), researchers introduce an approach that not only predicts what happens next, but also critically assesses whether an action was actually good. That’s exactly what makes this exciting for agentic systems: many models can extend a trajectory, but are much worse at distinguishing between “sounds plausible” and “was strategically sensible.” WAC therefore combines foresight with criticism—a pretty useful combination if you want to optimize policies, tools, or multi-agent workflows.
In practice, this means: better agents need not only better planning, but also a kind of built-in instinct to challenge themselves. This is especially relevant in the context of MCP workflows, delegation, and sub-agents, because otherwise errors just get automated elegantly. That’s less “autonomy” and more “scaled-up mistake with an interface.”
💼 Anthropic plans fund for workers threatened by AI
According to The Decoder, Anthropic wants to put five percent of its shares into a Public Benefit Fund that would also consider direct payments to workers under pressure from AI-driven disruption. That is notable because the company is addressing an issue that often only comes up once the layoffs are already underway: how the economic gains from AI are distributed.
What’s especially interesting is the signal this sends. Anthropic is positioning itself not just as a tech player, but also as an actor in the AI policy debate. At the same time, the fund remains under company control—so not a revolution, more a controlled attempt to reconcile responsibility with growth. That Elon Musk promptly suggested giving the shares directly to children fits the usual platform-theater section of the AI circus.
🏢 OfficeTown: Agents learn inside a living organization
With OfficeTown comes a generative simulation framework for professional organizations, where tasks, context, and feedback emerge from ongoing work. This matters because many current benchmarks measure isolated tasks—nice lab exercises, but far from real office life. OfficeTown tries to close that gap: agents should not just solve a task, but hold their own in an environment where knowledge about projects, documents, and feedback is distributed.
The bigger point: evolving agent intelligence needs environments that reflect change, dependencies, and social dynamics. If you want to know whether an agent is truly useful, it’s often not enough to test it on individual steps. In real life, work is rarely as tidy as a demo video. Unfortunately.
🔐 Passkeys and attestation: not a security gimmick, but context
heise explains in this background article why passkey attestation is not automatically a trap. That matters because security concepts are easy to misunderstand, especially when a viral post heats up the debate. In simple terms, attestation means a service can verify whether a passkey comes from a specific device or security element.
For you, that means: with passkeys, privacy, and device security, it’s worth looking at the concrete threat model rather than the buzzwords. Not every additional check is an attack on your data; sometimes it’s simply part of a more robust login system. If you judge security only by gut feeling, you often end up with the wrong gut feeling.
💸 MiniMax M3.1 Flash: cheaper, faster, but loves to hallucinate
MiniMax M3.1 Flash shows quite clearly where the market is headed right now: more performance per dollar. According to the report, the reasoning model scores 42 points on Artificial Analysis’ Intelligence Index, 13 more than its predecessor—at half the cost. That makes it interesting for anyone focused on LLM costs, inference, and production-adjacent use.
But: the hallucination rate is said to be 93 percent. That’s not a minor cosmetic flaw; it’s a reminder that cheap models are not automatically reliable models. For companies, that means: price matters, but so do evaluation, guardrails, and suitable use cases. Otherwise you save on the model and pay later in rework. A classic.
📱 WhatsApp opens calls to guests
With guest calls in WhatsApp, people without their own account will soon be able to join audio or video calls too—via browser on a PC. That sounds like a small product tweak, but it’s practically relevant: the barriers to meetings drop, especially when external contacts or one-off conversations are involved.
From a product perspective, it’s a good example of how platforms expand reach by reducing friction. For users, that’s convenient; for WhatsApp, of course, it’s also another step toward becoming a communication standard in everyday life. The real innovation here is not the tech, but the smoothness. And yes: surprisingly many tools fail right there.
🛠️ Tool tip of the day: test MCP workflows properly
If you work with agents, tools, and multi-step workflows, it’s worth looking at tools that handle MCP integration, debugging, and evaluation cleanly. Especially when delegating to sub-agents, it quickly becomes unclear where an error originated: in the model, the tool, or the orchestration.
My tip for today: use a setup that makes logs, tool calls, and intermediate states traceable before you switch agentic automation into production. It will save you quite a bit of detective work later. #
🛡️ AI agents as a weapon: the hacker attack with 100+ targets
The report from The Verge describes an attack in which an attacker allegedly used AI agents to compromise more than 100 websites. According to Reuters, victims included John Deere, SharkNinja, and Cummins. That is a rather sobering reminder that AI agents can become not only more productive, but also more dangerous when abused for cybercrime.
For security leaders, this means: automation on the attacker side scales effort, speed, and variation. Classic defense therefore has to rely more heavily on behavioral detection, access controls, and monitoring. For the AI industry, this is yet another sign that agentic systems must be evaluated not just for usefulness, but also for abuse potential. The bad news: apparently hackers read the same product announcements as everyone else.
Don’t want to miss any news? Subscribe to the newsletter