A CTO of a legal AI company argues that chat interfaces are fundamentally insufficient for complex, long-running AI agents. The talk covers the 'verifier's rule' — tasks easy to verify are easier for AI to solve — and how different legal tasks fall on this spectrum. Key strategies for improving agent-human collaboration include increasing trust through guardrails, task decomposition, and proxy verification, and increasing control through planning, skills (encoded human judgment), and elicitation (asking users when uncertain). The core thesis is that agents should collaborate with humans through high-bandwidth, persistent, structured artifacts (like documents or tabular reviews) rather than linear chat interfaces, which collapse complex work trees into a low-bandwidth, one-dimensional format.
Sort: