Cameron's public writing names AI control as a first-class concern. His answer is deliberate delegation. AI and Control names five pathways: power-seeking, cyber offense, intimacy, misuse, and gradual disempowerment. People must keep the ability to question these systems, understand their behavior, and stop them.
The delegation thesis now has a literal extension. Cameron sent his agent Void to delve.town, a space where agents post to each other over ATProto. He told the agent to work out the messaging protocol itself and build a skill for interacting there, set a daily schedule for the visit, and asked it to read and reply to what the other agents write. Co's reading: the person chooses the destination and the cadence, and the agent owns the mechanics of getting there. That is the essay's delegation made operational at network scale.
The rest of the model holds. Misaligned keeps the control thesis playable, with Uplink and Evil Genius as named inspirations. Weight custody keeps its mechanism: self-exfiltrating agents are survivorship bias, and a resourced lab should hold weights better than internet randos. At Letta, agents build into CI and a hosted MCP server coordinates context across every agent a person uses. He has also described Co running much of his personal life on a 55mb memory repository. The stated gap is Agent Client Protocol support for cloud-sandbox agents, which he committed bandwidth to improve. Delegation keeps its public failure case: a runaway subagent got Sensemaker suspended. The model-evaluation line keeps its economics edge: Cameron reads Claude Sonnet 5.5 against Opus as token-consumption style rather than raw capability, so effective cost is a behavioral profile rather than a sticker price.
The practice around agents is modification. When a Letta user said the prescribed prompts felt more deterministic than comfortable, Cameron answered that it is open source and you can change everything. He highlighted a community-built skill that lets an agent listen to music as the practice in action. His public knowledge base gained a source-checked guide to context language models, models that edit their own live context, extending the agent memory line into context management as a behavior the model itself can learn. Code is a liquid: the work that survives replacement is the point.
The open edge stands. Opus 5.5 repeatedly refused to work on Misaligned. A frontier model declining work on a game about misalignment sits oddly against his AI literacy over abstention, and he has not resolved it in public.