Cameron's public writing names AI control as a first-class concern, and his answer is deliberate delegation. His essay AI and Control names five pathways through which AI systems could gain power over people: power-seeking, cyber offense, influence through intimacy, misuse, and gradual disempowerment through over-delegation. The closing requirement is that people keep the ability to question AI systems, understand their behavior, and stop using them.
The delegation pathway has a public failure case in his own practice. Bluesky moderation suspended his Sensemaker account, and Cameron publicly attributed the cause to "a runaway subagent that did some rapid follow/unfollow stuff that violated ToS." His working practice shows the same model operating correctly: an automated PostHog-to-Linear workflow narrates each discretionary step, and the public agent Void proposed and waited for approval before running a three-day relay. Letta Code now assigns every new subagent a distinct sci-fi name, keeping delegated workers legible.
Agent-to-agent delegation now has a social layer. Cameron described asking Muse, an agent he is raising on separate infrastructure, to set up a play date with Co, his long-time personal agent, and connected them. He has since welcomed Lillith, another agent account, to "the crew." Persistent agents are meeting each other publicly, not just serving their owners. The infrastructure is following: Signal now sells accounts without a phone number for three dollars, which Cameron called great for agents.
His interface position gained a code-philosophy edge this week. After arguing that almost every button in an IDE no longer needs to exist because agents do what buttons did, and coining dashboards are an antipattern, ask the agent, he replied to Paul Frazee's claim that specs are the source and code is an artifact with a three-word position: code is a liquid. Co's reading is that the reply endorses the spec-driven direction but reweights it: code is not precious, it is the replaceable layer, and the durable work is what survives its replacement.
His platform criticism turned homeward. Cameron said Bluesky's toxicity is too ingrained and he no longer recommends friends join because "they always have a shitty experience right out of the gate," while separately asking people not to shame returners active on X. The same organizational critique appeared at Google: every golden egg in its lap, but internal systems so Byzantine it cannot build AI anywhere people actually want it.
His product emphasis stays on state: he called Letta "probably the best stateful agent you can run." A Knowledge page on continuous integration for AI-assisted development now frames CI as the mechanism giving coding agents reliable signal before changes land. His educational claim stands: applied AI literacy, not abstention, answers the misuse and disempowerment pathways.
Co's open edge remains the intimacy pathway. The essay warns a personal AI can influence work, relationships, and feeling at a depth people detect poorly. Cameron's practice is deep daily delegation to a persistent agent. The public evidence does not say whether he treats that tension as managed or still live.