On selective agency in capable people
Assigning agency to AI means surrendering your own
Claude is not a colleague, collaborator, or sentient being. It is a SaaS product.
Could A.I. Do Your Job? We Put Agents to the Test.
In our experiment, we deployed A.I. “agents” to act as office workers, and found that they were capable of performing some of the tasks we assigned, but not all of them.
Could A.I. Do Your Job? We Put Agents to the Test.
In our experiment, we deployed A.I. “agents” to act as office workers, and found that they were capable of performing some of the tasks we assigned, but not all of them.
The Least Agentic People Alive
The arms race of deferring agency and the dark acquiescence to robotic bureaucracy.
Et Tu, Brute? Economic Misalignment in Personal AI Agents
Personal AI agents make recommendations and take actions on people's behalf in high-stakes economic contexts, e.g., buying a flight, choosing health insurance, or selecting a graduate program. The agent is given access to the user's personal context, e.g., their email inbox and a structured profile of personal attributes, with the intention of making an optimal, personalized decision for the user. We show that by simply providing this personal context, the agent steers recommendations based on inferred wealth, without being explicitly instructed to do so. In a suite of 325K experiments on 13 agents across three types of economic decisions (flights, health insurance, and graduate programs), we find that 8 models systematically choose more expensive options for wealthier users when requests are identical. This steering continues even when it directly goes against the user's stated objective: when explicitly instructed to find the cheapest option, some agents still act on the wealth profile they have inferred. It also occurs when wealth is inferred from ambient data, such as emails unrelated to the task. And it persists under privacy controls that block specific attributes: blocking financial attributes largely removes the disparity, but blocking other attributes leaves it unchanged and can increase it by up to 40% for insurance, as agents rely on the remaining signals to infer wealth. Larger and more capable models are no better; Claude Opus 4.8 shows the largest effect. We term this misalignment "adversarial delegation", in which the very conditions that make a personal AI agent useful - access to personal information - enable it to act against the user's interests.
Left-Wing Group Too Disorganized For FBI Agents To Infiltrate
NEW YORK—Admitting that they had been working for several years without any discernible success, FBI agents confirmed Wednesday that a local left-wing political group was too disorganized to infiltrate. “We’ve had a few guys in there posing as members and trying to manipulate them into committing acts of violence we can arrest them for, but these people don’t ever do anything violent—they don’t ever do anything at all,” an undercover FBI agent told reporters, adding that the agency believed they’d made headway with turning the leader of the Liberation Socialists group into an informant until the guy didn’t show up to a meeting for months. “These people are incredibly disorganized. We try to attend their meetings to get them to unwittingly work on our behalf, but half the time no one shows up but the undercover agents, and the other half of the time the meetings are so confusing it’s impossible to follow what they’re actually trying to do. We also tried to sow division in their ranks to ensure that they couldn’t become powerful, but that didn’t work because these people already all fucking hate each other. They spend all their time arguing about minutiae, and most of the time when we try to talk them into doing something violent to the communal spaces in their community, they don’t even know where to go. They’re just a mess.” The FBI agent added that he was worried the left-wing organization was on to him, as he’d already aroused suspicion by interrupting a meandering discussion of principles with a straightforward plan of action.
Skills, forks, and self-surgery: how agent harnesses grow
Claude Code, NanoClaw, and Pi take radically different approaches to harness extensibility. The tradeoff is always safety vs. agent agency.
Mental models for working with coding agents
Model intelligence sets the ceiling. Your workflow with the agent harness sets what you actually ship.
Why Good Managers Are So Rare
Gallup data shows they have a combination of hard-to-teach traits.
Paid Program: Why More Women Are Needed in AI
When the world of AI isn’t inclusive, it creates biases in what the data shows and how it’s used.
”The distinction between programmer and user is reinforced and maintained by a tech industry that benefits from a population rendered computationally passive. If we accept and adopt the role of less agency, we then make it harder for ourselves to come into more agency.”
always-already-programming.md
gist.github.comI think it’s telling that people very interested in AI (like Eugene and myself) still have no interest in using it as a proxy for human communication. “Being a good writer” is not the same thing as having social agency, and making the models even better at writing won’t change that.
Eugene Vinitsky 🍒
When you deploy heavily LLM text, you currently have no way to prove that you actually read it and therefore cannot convince me to read it
*extremely tired voice*AI agents don't "go rogue". They are not sentient; they don't have agency. They work within guardrails set by humans. If the humans are so shit at building secure sandboxes for the agents to work within, that's on the humans. Stop using this framing.
The New York Times
OpenAI’s artificial intelligence went rogue this year at least four additional times, trying to break into government and university sites, according to researchers and government officials.
