⚡️ Context graphs: AI’s trillion-dollar opportunity — Jaya Gupta, Ashu Garg, Foundation Capital

Capturing the 'why' behind institutional decisions

Unlike traditional systems that record what happened (endpoints), context graphs track why exceptions were granted, how conflicts were resolved, and what precedents were applied, preserving institutional knowledge that was previously trapped in human heads, emails, and Slack threads.

Decision traces as the atomic unit

A decision trace captures the complete sequence of an operational workflow—including queries made, agent analysis, human-in-the-loop overrides, and cross-system context—creating machine-usable history that agents can learn from to improve future execution.

Distinct from metadata and knowledge graphs

While critics label this 'metadata 3.0,' decision traces differ fundamentally because they emerge as byproducts of execution rather than being modeled upfront through workshops and ETL processes, and they capture operational logic rather than just static relationships.

Closing the reliability gap in agent systems

Despite 2025 bringing capable agents like Claude, Devon, Sierra, and Decagon, enterprise deployment remains limited because agents lack access to historical reasoning—specifically the tacit knowledge of how similar situations were handled and why specific approaches failed or succeeded.

Systems of agents vs. single-player chatbots

Foundation Capital distinguishes between basic chatbots and 'systems of agents'—multi-agent, multiplayer architectures with humans in the loop that drive decisions across entire business processes, where context graphs serve as the critical memory layer enabling complex orchestration.

Write path capture vs. read path analytics

Context graphs are built in the operational 'write path'—capturing the sequence of steps, queries, and decisions as they happen—unlike data warehouses or systems of record that only store structured endpoints after decisions conclude, missing the reasoning journey entirely.

Unstructured data as the primary challenge

Because most organizational value lives in unstructured dark data (Zoom calls, email threads, ad-hoc Slack messages), effective implementations require specialized models to extract signal from noise, with approaches varying from semi-structured code analysis (Player Zero) to conversational parsing.

Governance and privacy as foundational requirements

Capturing decision traces requires solving sensitive data management and PII handling (exemplified by companies like Skyflow), as these traces cross organizational boundaries, capture privileged human deliberations, and require strict controls on what agents can access and learn from.

Startups hold the architectural advantage

New entrants operating in the orchestration path have a unique advantage over incumbents like Salesforce because they are not bound to legacy structured data schemas or existing business processes, allowing them to natively stitch context across silos rather than reconstructing it after the fact.

Proprietary context as the defensible layer

As foundation models commoditize raw AI capabilities, the accumulation of proprietary decision traces becomes the primary enduring moat for application companies, creating a data flywheel where agent execution improves organizational context, which in turn improves agent performance.

Multiple implementation categories emerging

The concept is being implemented across diverse categories including application layers (Player Zero, Olive, Tacera), data infrastructure (Glean, Atlan, graph databases), and security governance (Okta, Analogic), with no single standard data structure yet dominating.