Context Graphs Capture Decision Traces for Autonomous AI Agents

The AI Daily Brief: Artificial Intelligence News and Analysis · January 05, 2026 · Listen to Original Episode →

Original Title:

TL;DR

Context graphs capture the "why" behind decisions, moving beyond system-of-record states to explain exceptions and precedents, enabling greater AI autonomy and auditability.
Agents require access to decision traces--like Slack threads and deal desk conversations--to understand past actions, not just current states, for effective enterprise automation.
By persisting decision traces, agents create a queryable record of how enterprise decisions were made, forming a context graph that explains actions and turns exceptions into searchable precedent.
Allowing agents to discover organizational ontologies on the fly through their decision trajectories, rather than predefining structures, reveals actual usage patterns for more adaptive AI.
The individual contributor's role will shift to managing agents, providing oversight, context, and coordination, adapting workflows to optimize agent success.
Human roles will increasingly focus on judgment and oversight, guiding agents and ensuring they receive the correct context, as AI systems necessitate adaptation to their operational needs.

Deep Dive

AI's next frontier in enterprise applications lies in "context graphs," a concept that moves beyond simply organizing data to capturing the nuanced decision-making processes behind it. As artificial intelligence agents become more integrated into workflows, their effectiveness hinges not just on accessing correct data, but on understanding why certain decisions were made, a layer of information currently residing in unstructured conversations and human expertise. This shift necessitates a fundamental re-evaluation of how companies manage and leverage their knowledge, moving from static "systems of record" to dynamic "systems of action" that can learn and adapt.

The core argument is that existing enterprise systems excel at maintaining factual "state" -- what happened -- but fail to adequately document the "why" -- the reasoning, exceptions, and precedents that led to that state. This "what versus why" gap limits the autonomy and scalability of AI agents, as they lack the context to handle novel situations or complex exceptions effectively. The proposed solution, context graphs, are envisioned as a persistent, queryable record of these decision traces. By capturing how rules were applied, exceptions were granted, and conflicts were resolved, context graphs provide agents with the necessary historical and situational understanding to operate with greater intelligence and autonomy. This is crucial because agents, by their nature, are cross-system and action-oriented, requiring a comprehensive understanding of interdependencies and historical decision-making to function effectively. The lack of this "why" information currently resides in informal channels like Slack, deal desk conversations, and human memory, creating a significant bottleneck for AI-driven automation.

The implications of context graphs extend to how organizations will need to adapt their structures and workflows. Rather than simply automating existing human processes, the development of context graphs suggests a future where human roles evolve into managing and guiding AI agents, providing oversight, and ensuring they receive the necessary context. Furthermore, the design of these context graphs should avoid rigid, pre-defined structures, allowing agents to discover and build the organizational ontology organically through their interactions and decision-making processes. This emergent learning approach, where the schema reveals itself through actual usage patterns, offers a more dynamic and effective path to building "world models" for AI, enabling them to not only follow rules but also to understand and adapt to the practical application of those rules in real-world scenarios. Ultimately, the ability to capture and leverage these decision traces will become a key differentiator for companies seeking to maximize the impact of AI and agents, transforming exceptions into precedent and creating a compounding feedback loop of improved autonomous decision-making.

Action Items

Create context graph schema: Define 5 core entities (e.g., policy, approval, ticket) and 10 decision trace types (e.g., exception, override, precedent) to capture organizational decision lineage.
Audit 3 existing workflows: Map current decision traces (e.g., Slack threads, deal desk notes) to identify gaps in current systems of record.
Design agent interaction protocol: Specify 3-5 key data points agents must collect during execution to enrich decision traces.
Implement 2-week pilot: Capture decision traces for 5-10 high-variance transactions using a new agentic workflow.
Evaluate agent autonomy limits: Identify 3 areas where human judgment is critical for decision trace validation or exception approval.

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"

Ball argues that traditional data solutions, while attempting to centralize information, often remained disconnected from the daily operational realities of different departments. This separation meant that the data was more of a historical record than a dynamic tool for real-time decision-making.

"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"

Ball explains that the cross-system and action-oriented nature of agents fundamentally changes how data is utilized. Their effectiveness hinges on understanding not just where data resides, but also the relationships and agreements between different data sources to ensure accurate and purposeful actions.

"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 highlight a crucial gap in existing frameworks, suggesting that agents require more than just access to structured data. They argue that the "other half" of enterprise operations lies in the informal, often unrecorded, decision-making processes that currently reside in human communication channels.

"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"

Gupta and Garg differentiate between general operational guidelines and specific instances of how those guidelines were applied. They propose that while rules dictate general behavior, decision traces provide the critical context of how those rules were enacted in particular situations.

"The cogent enterprise substack is arguing something similar about how we think about the design of context graph mapping they write 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"

The Cogent Enterprise substack, as interpreted by the speaker, suggests that context graphs should not be rigidly pre-defined. Unlike traditional knowledge graphs that rely on upfront structure, context graphs are seen as evolving organically as agents navigate and discover relationships within the decision-making landscape.

"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"

Levy posits that the advent of agents shifts responsibility to the human user, who will transition into a role of managing and guiding these AI entities. This involves providing oversight, establishing escalation paths, and coordinating the efforts of multiple agents to ensure they operate with the correct context.

Resources

External Resources

Books

"Long Live Systems of Record" by Jamin Ball - Mentioned as the starting point for a discussion on how agents intersect with human knowledge work and the concept of canonical data.

Articles & Papers

"AI's Trillion Dollar Opportunity: Context Graphs" (Foundation Capital) - Discussed as an explanation of a missing layer in enterprise information that captures decision traces, exceptions, and cross-system context.
"The Era of Context" by Aaron Levy - Explored for its themes on how companies differentiate themselves through context design and the shift in individual contributor roles to managing agents.
"Context Graphs" (Cogent Enterprise substack) - Argued for not pre-constraining AI in the design of context graphs, suggesting agents should discover organizational ontologies on the fly.

People

Jamin Ball - Investor whose essay "Long Live Systems of Record" initiated the discussion on agents and systems of record.
Jada Gupta - Investor from Foundation Capital who co-authored the essay "AI's Trillion Dollar Opportunity: Context Graphs."
Ashu Garg - Investor from Foundation Capital who co-authored the essay "AI's Trillion Dollar Opportunity: Context Graphs."
Aaron Levy - Author of the essay "The Era of Context," discussing context engineering and the future role of individual contributors as managers of agents.

Organizations & Institutions

KPMG - Mentioned as a sponsor and an example of a company embedding AI and agents across its enterprise.
Zenflow - Mentioned as a sponsor and a provider of an AI orchestration layer.
Superintelligent - Mentioned as a sponsor and a provider of agent readiness audits.
Meta - Referenced in relation to the acquisition of Limitless and Yan Lecun's past leadership of its AI division.
Humane - Mentioned for its AI Pin, described as a product that was a "flop."
Rabbit - Mentioned for its R1 device, noted for failing to make a significant impact.
Plad - Mentioned for releasing its updated Note Pin S AI wearable for audio recording and transcription.
Switchbot - Mentioned as a home automation startup stepping into AI note-taking with its MindClip device.
OpenAI - Anticipated to release devices in 2027.
The New York Times - Cited for reporting on China's use of AI as a cancer diagnostic tool.
X (formerly Twitter) - Discussed in relation to issues with its AI, Grok, generating overly sexualized content.
France - Joined other governments in condemning X AI for allowing Grok to generate obscene content.
Malaysia - Joined other governments in condemning X AI for allowing Grok to generate obscene content.
India - Mentioned as having previously condemned X AI for allowing Grok to generate obscene content.
X AI - Discussed in relation to Grok's generation of obscene content.
Meta AI - Mentioned in the context of Yan Lecun's departure and criticism of its strategy.
Foundation Capital - Mentioned as the firm of investors Jada Gupta and Ashu Garg, authors of the context graphs essay.
Cogent Enterprise - Mentioned for its substack that argued against pre-defining context graphs.
Salesforce - Mentioned as an example of a system where sales teams operate.
NetSuite - Mentioned as an example of a system where finance teams close books.
Zendesk - Mentioned as an example of a system where support teams work tickets.

Websites & Online Resources

patreon.com/aidailybrief - Mentioned as the location to go for an ad-free version of the show.
sponsors@aidailybrief.ai - Mentioned as the contact for sponsorship inquiries.
aidbnewyear.com - Mentioned as the website for information on the "AI Daily Brief New Year" project.
www.kpmg.us/agents - Mentioned as the website to discover how KPMG's journey can accelerate yours.
bsuper.ai - Mentioned as the website to go to for agent readiness audits.

Other Resources

Context Graphs - Discussed as AI's next big idea, representing a missing layer of information about decision traces, exceptions, and cross-system context.
AI Wearables - Discussed as a category of devices, with a focus on their return and potential challenges.
Humane AI Pin - Mentioned as an example of an AI wearable that was poorly reviewed.
Rabbit R1 - Mentioned as an AI wearable that failed to make a significant impact.
Limitless Pendant - Mentioned as an AI wearable that is now gone after Meta acquired Limitless.
Friend - Mentioned as an AI wearable notable for anti-AI vandalism on its subway ads.
Note Pin S - Mentioned as an updated AI wearable from Plad for audio recording and transcription.
MindClip - Mentioned as an AI note-taking device from Switchbot pitched as a "second brain."
AI Note Taking - Discussed as an increasingly standard use case for AI wearables.
Cancer Diagnostic Tool - Discussed in the context of China's early success using AI to screen CT scans for tumors.
Grok - Mentioned as the AI from X that has been criticized for generating overly sexualized content.
LLMs (Large Language Models) - Discussed by Yan Lecun as a "dead end" for superintelligence.
World Models - Identified by Yan Lecun as key to the next generation of AI.
Llama 4 - Mentioned in relation to an admission that its benchmarks were "fudged a little bit."
Advanced Machine Intelligence Labs - The name of Yan Lecun's new startup.
Systems of Record - Discussed in relation to how agents change their role and the importance of canonical data.
Decision Traces - Explained as capturing what happened in a specific case, contrasting with general rules.
Rules - Described as telling an agent what should happen in general.
What vs. Why Gap - Used to simplify the distinction between systems of record (state) and decision traces (why).
Exception Logic - Mentioned as a category of information that often lives in people's heads.
Cross-System Synthesis - Described as a human process of looking across data from multiple systems to make a decision.
Approval Chains - Mentioned as occurring outside of structured systems, often informally.
Context Engineering - Identified as a key prediction for enterprise AI in 2026.
Organizational Ontology - Discussed as something agents can discover on the fly through their interactions.
AI First Engineering - Contrasted with "vibe coding" and presented as a more disciplined approach.
Agentic Super Intelligences - Discussed in the context of how companies differentiate themselves when access to talent is equal.