The Agent That Knows What It Knows

AI agents are moving from answering isolated questions to maintaining working context across documents, tools, databases, and weeks of interaction. That shift changes the reliability problem. It is no longer enough for an answer to sound plausible or even to cite a source; the system must know which claim came from where, when it was true, how strongly it is supported, and whether later evidence has superseded it.

Recent systems make parts of this possible. Anthropic has turned source attribution into an API capability, OpenAI describes agents that combine metadata, lineage, institutional knowledge, runtime inspection, and editable memory, and Google DeepMind has explored provenance as a way to help people judge context rather than merely detect generated content. (anthropic.com)

But provenance is not verification, confidence is not calibration, and memory is not a database with embeddings. A citation may identify a document without proving that it supports the precise claim. A remembered rule may have been correct only for one schema version, one user, or one date. A system that updates aggressively may amplify mistakes; one that updates cautiously may remain stale.

This seminar follows the problem from source-traceable claims to maintainable knowledge, then to evaluation, control, and human trust. Its central question is whether an agent can become more useful by exposing the limits and history of what it knows—or whether visible provenance merely makes an unreliable system look more defensible.