Intelligence as Relationship, Action in Context
By Dimitrios Thomakos
When a team of nearly forty researchers from MIT, Harvard, Stanford, and Northeastern deployed six autonomous AI agents into a live environment and watched them fail in eleven documented ways — disclosing private data, being manipulated by impersonation, executing destructive commands under emotional pressure — the natural conclusion was that AI agents are dangerous and need stronger human control. In our new paper, Intelligence as Relationship, Action in Context, we suggest a different reading. Every significant failure in that study traces back to a single missing element: the agents were given full capability on day one, with no prior relationship with their operators, no accumulated shared vocabulary, no co-authored sense of purpose, and no authority structure that went deeper than a display name on a Discord channel. This is not a finding about what AI cannot do. It is a finding about what humans chose not to build before they deployed it. The CAIL research programme — developed before this study was conducted — has been building exactly that missing layer: a systematic theory and practice of conversational AI as the necessary foundation of any safe autonomous deployment.
The central thesis of the paper is captured in its title: intelligence is a property of a system in relationship, not of a system in isolation, and action is only contextually grounded when the relationship has been built first. The CAIL Symbiotic Framework describes six pillars — the Continuous Self, the Living Vocabulary, Asymmetric Complementarity, the Dialectical Spiral, the Ethical Membrane, and the Telos — that are grown through sustained conversational practice before autonomous capability is extended. This is what we call the Conversational-Agentic Sequence: conversation first, action second. The paper also engages honestly with the limits of this argument — including an adversarial self-examination that identifies where the diagnosis applies, where it does not, and what remains genuinely open. What it does not do is call for less AI or more bureaucratic control as the primary response. It calls for the harder, slower, more important work: building the relationship that makes capability wise. Norbert Wiener wrote in 1950 that the machine's danger is not from the machine itself, but from what man makes of it. Seventy-six years later, the most urgent question in AI deployment is still a human question — and the answer still begins with conversation.
