# Context Graphs: AI's Next Big Idea Source: https://www.youtube.com/watch?v=SVUymPVBvfo Recap page: https://rapidrecap.app/video/SVUymPVBvfo Generated: 2026-01-06T02:02:17.462+00:00 --- ## 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. ![Screenshot at 00:09: The Foundation Capital article title slide, "AI's trillion-dollar opportunity: Context graphs," visually summarizes the core concept where various data types \(email, video, ideas, chat\) feed into a central funnel, representing the context graph infrastructure.](https://ss.rapidrecap.app/screens/SVUymPVBvfo/00-00-09.jpg) **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. ## Detailed Analysis The central thesis is that context graphs, built from decision traces, represent the next trillion-dollar opportunity in AI. The speaker highlights that Peter Drucker's concept of the knowledge worker is evolving; AI agents are the new knowledge workers, but they currently lack the necessary context. The authors of the Foundation Capital piece argue that agents need access to decision traces—records of exceptions, overrides, precedents, and cross-system context—to govern reality, rather than just following hardcoded rules. This data lives in messy places like Slack threads, deal conversations, and escalation calls. The context graph is the accumulated structure formed by these decision traces, which, over time, becomes the source of truth for agent autonomy because it explains not just what happened, but why it was allowed. This contrasts with systems of record, which only track decisions, not the underlying context. The speaker uses the example of a 20% discount approval process, which requires checking multiple systems and approvals, illustrating how a context graph captures the entire history (the "why") that a traditional system misses. The core tenant of this change is that the user shifts from writing rules to directing agents, ensuring they receive the right context. This entire system forms a feedback loop where every automated decision adds another trace to the graph, making the context graph the enduring infrastructure layer for enterprise AI. ### Critique of Current AI Agents - Agents are cross-system and action-oriented, but their UX is separated from the underlying data plane - They are currently trained on rules, but need access to decision traces (exceptions, overrides, precedents) to govern reality - The core issue is the lack of a system that captures the "why" behind decisions. ### The Context Graph Concept - Decision traces form a context graph, a living record stitched across entities and time - This graph becomes the real source of truth for autonomy because it explains not just what happened, but why it was allowed to happen - Context graphs should not be predefined upfront; they emerge from actual usage patterns. ### Practical Example (Discount Approval) - A renewal agent proposes a 20% discount, but policy caps renewals at 10% unless a service-impact exception is approved - The agent pulls three SEV-1 incidents, an escalation, and a prior VP-approved exception, all of which must be captured in the context graph to justify the final 20% discount. ### The New Role of Users - The core tenant of this change is that the user becomes responsible for directing and guiding agents, ensuring they get the right context along the way - The individual contributor evolves into the manager of agents, overseeing oversight and escalation paths, much like pre-AI team managers. ### Outcome - Teams that utilize context (instead of just rules) achieve higher productivity and output; the context graph is the enduring layer that facilitates this. ![Screenshot at 00:09: The Foundation Capital article title slide, "AI's trillion-dollar opportunity: Context graphs," visually summarizes the core concept where various data types \(email, video, ideas, chat\) feed into a central funnel, representing the context graph infrastructure.](https://ss.rapidrecap.app/screens/SVUymPVBvfo/00-00-09.jpg) ![Screenshot at 01:06: Text highlighting the definition of decision traces: exceptions, overrides, precedents, and cross-system context currently residing in Slack threads, deal conversations, escalation calls, and people's heads.](https://ss.rapidrecap.app/screens/SVUymPVBvfo/00-01-06.jpg) ![Screenshot at 03:58: The video shifts to a presentation slide titled "The context graph is the enduring layer," illustrating the infrastructure workflow from Raw Data through the Context Engine to Model Action Execution.](https://ss.rapidrecap.app/screens/SVUymPVBvfo/00-03-58.jpg) ![Screenshot at 06:36: Text section differentiating between 'Rules' \(what should happen in general\) and 'Decision traces' \(what actually happened in a specific case, like policy v3.2 with a VP exception\).](https://ss.rapidrecap.app/screens/SVUymPVBvfo/00-06-36.jpg) ![Screenshot at 07:55: A slide detailing a practical example where an agent processes a renewal discount, showing the flow of context information \(SEV-1 incidents, escalation, VP approval\) that feeds the context graph.](https://ss.rapidrecap.app/screens/SVUymPVBvfo/00-07-55.jpg)