Context Graphs: AI's Next Big Idea

Quick Overview

Context graphs, which capture the full context of agent decisions via decision traces, represent the trillion-dollar opportunity in AI, shifting the focus from just rules to understanding the "why" behind actions, as demonstrated by the Foundation Capital essay and the inherent complexity of real-world enterprise workflows.

Key Points: Context graphs capture decision traces—exceptions, overrides, precedents, and cross-system context—which are the missing layer in current enterprise AI systems. The core distinction that matters is between static 'Rules' (what should happen) and dynamic 'Decision Traces' (what actually happened, including exceptions and precedents). Context graphs allow agents to become 'Informed Walkers' that discover organizational ontology on the fly, learning how entities relate through actual use rather than predetermined schemas. The feedback loop of captured, searchable decision traces creates a 'living record' that becomes the real source of truth for autonomy, explaining not just what happened, but why it was allowed. The shift implies that the user's responsibility moves from writing rules to directing and guiding agents, ensuring they get the right context along the way. A concrete example showed a 20% discount approval routing through multiple systems (CRM, Finance) where the context graph captures the entire exception trail, unlike traditional systems. The authors argue that context engineering, rather than the model itself, is the moat and core infrastructure for future trillion-dollar AI platforms.

Context: The video discusses the concept of "Context Graphs" as proposed in an essay by Jaya Gupta and Ashu Garg from Foundation Capital, arguing it is the next major area for AI investment beyond large language models. The discussion centers on how current AI agents struggle with real-world enterprise complexity because they rely on rules or siloed data, leading to issues like inconsistent decision-making and a lack of historical context for exceptions.