agent reliability

Your AI agent's outputs look composed. Its internal state is desperate. Anthropic's emotion vectors research reveals a second axis of agent drift that output evals can't catch.

Autonomous AI agents degrade over long-horizon tasks through accumulated context pollution, not single failures. What agent drift is, why it happens, and how to architect against it.