Context Graphs: Capturing Decision Traces for Scalable AI Autonomy

Key Quotes

"The problem is that most of this lived downstream of the operational world the sales team still lived in salesforce the finance team still closed the books in netsuite the support team still worked tickets in zendesk the warehouse for lakehouse was the retrospective mirror not the transactional front door."

This quote from Jamin Ball highlights a fundamental disconnect in traditional enterprise data management. Ball argues that operational systems remain siloed within departments, and data warehouses or lakes, while retrospective, do not actively influence or guide daily operations. This separation means that even with advanced data infrastructure, the actual transactional front doors of businesses remain disconnected from a unified, operational view of data.

"Agents are inherently cross system meaning they don't live solely within one of those functions and they are action oriented they are not just trying to gather information they are trying to make use of that information to do things that combination he writes means agents are only as good as their understanding of which system owns which truth and what the contract is between those truths under the hood something still has to say this is the canonical customer record or this is the legally binding contract term or this number is the one we report to wall street that something might be a traditional system of record it might be a warehouse back semantic layer or it might be a new class of data control plane product but it is absolutely not going away."

Ball explains that agents, by their nature, operate across different systems and are designed to act on information, not just collect it. This capability makes their effectiveness dependent on understanding which system holds the definitive truth for specific data points and how these truths interrelate. Ball emphasizes that a central authority, whether a traditional system of record or a newer data control plane, is essential to establish and maintain these canonical truths for agents to function reliably.

"Ball's framing assumes the data agents need already lives somewhere and agents just need better access to it plus better governance semantic contracts and explicit rules about which definition wins for which purpose that's half the picture the other half is the missing layer that actually runs enterprises the decision traces the exceptions overrides precedents and cross system context that currently lives in slack threads deal desk conversations escalation calls and people's heads."

Gupta and Garg introduce the concept of "decision traces" as a critical missing layer in enterprise AI, complementing the idea of systems of record. They argue that while systems of record manage the "what" (e.g., a deal's discount), they fail to capture the "why" -- the specific context, exceptions, and precedents that led to a decision. This crucial "why" information, often residing in informal communication channels and human knowledge, is essential for agents to operate with true autonomy.

"The distinction the authors say is between rules and decision traces rules they say tell an agent what should happen in general whereas decision traces capture what happened in this specific case i think for our purposes on this episode an even simpler way to understand it is the what versus why gap and the simple idea here is that while systems of record are good at state i e this particular deal closed at a 20 discount they are bad at decision lineage why a 20 discount was allowed this time as the authors point out those decision traces i e the why lives in slack and dms and in meetings and in human heads limiting how much autonomy can then scale as the authors put it agents don't just need rules they need access to the decision traces that show how rules were applied in the past where exceptions were granted how conflicts were resolved who approved what and which precedents actually govern reality."

Gupta and Garg clarify the difference between rules and decision traces, framing it as the "what versus why" gap. They explain that rules provide general guidance, while decision traces document the specific rationale, exceptions, and approvals behind individual actions. This "why" information, often found in informal communications and human memory, is crucial for agents to understand past decision-making processes and scale autonomy effectively.

"The most counter intuitive development we shouldn't pre define these context graphs traditional knowledge graphs fail because they require pre defining structure upfront context graphs invert this completely modern agents act as informed walkers through your decision landscape as an agent solves a problem traversing through apis querying documentation reviewing past tickets it discovers the organizational ontology on the fly it learns which entities actually matter and how they genuinely relate through use not through a manual schema someone designed in a workshop each trajectory leaves a trace which systems were touched together which data points co occurred in decision chains how conflicts were resolved accumulate thousands of these walks and something remarkable emerges the organizational schema reveals itself from actual usage patterns rather than predetermined assumptions these become world models not just retrieval systems."

The Cogent Enterprise substack argues against pre-defining context graphs, contrasting them with traditional knowledge graphs that require upfront structure. They propose that modern agents can discover an organization's ontology organically by "walking" through decision landscapes, interacting with APIs, and reviewing data. This process, driven by actual usage patterns rather than manual schema design, allows the organizational schema to reveal itself, forming "world models" rather than mere retrieval systems.

"Designing our systems to get agents access to that data and ensuring that all of our agents can interoperate on that data is going to be incredibly important further companies will have to drive a substantial amount of change management to make this all work we imagined that ai systems would adapt to how we work but it turns out due to their extreme power and inherent limitations we will instead adapt to how they work this means we will have to optimize our organizations and workflows to best enable context for agents to be successful the core tenet of this change is that the user is responsible now for directing and guiding agents on how to do their work ensuring it gets the right context along the way in essence he writes the individual contributor of today becomes the manager of agents in the future their new responsibilities will be providing the oversight and escalation paths a meaningful amount of coordination throughout the work that the agents are doing and shepherding work between the various agents just like managers of teams in the pre ai era."

Aaron Levy emphasizes the critical role of "context engineering" and the need for organizational adaptation to AI. Levy posits that companies must design systems for agent data access and interoperability, which will necessitate significant change management. He suggests that humans will adapt to how AI systems work, optimizing workflows to provide agents with the necessary context, and that individual contributors will evolve into "managers of agents," responsible for oversight, coordination, and guiding AI tasks.