Every company needs a Cassandra
Useful framing for agents as organizational cognitive functions: not only workers that do tasks, but systems that preserve dissent, memory, and counterfactual thinking.
Links: Original source · Shared link · Related link
Logged at IST: 2026-08-10 10:18 IST
What it is: Sunil Pai’s essay proposing Cassandra, a background AI agent for the socially expensive work of organizational dissent.
Gist: Pai’s core point is that companies say they want disagreement, but repeated disagreement costs humans status, trust, and invitation rights. A useful agent could absorb some of that cost by watching discussions, keeping a private model of assumptions and past decisions, and speaking only when the expected value of interruption is high.
The important distinction is that Cassandra should not be a generic devil’s advocate. She needs independent context, memory, and restraint: competitor docs, support tickets, failed deals, postmortems, previous predictions, and enough distance from the company line to notice when consensus is missing something. The failure mode is also clear: humans cannot outsource dissent to the bot. The agent should start better disagreements, not replace them.
Newsletter angle: Good framing for agents as missing organizational cognitive functions, not just extra labor. The interesting question is which useful ways of thinking are too awkward, tiring, or politically costly for a person to perform continuously.