1 Here’s the core idea: most of our current systems capture what happened, but not why it happened.
- Why did this deal need to be escalated to legal review?
- Why did we pick Providence, RI for our next retail store?
- Why did we decide to discontinue product [X]?
That reasoning the decision traces, the exceptions, the precedents lives scattered across Slack, work calls, and inside people’s heads. It’s insider knowledge that builds up as employees gain experience and resets every time someone leaves.
A context graph is meant to capture all of that systematically. Not just the final state, but the full sequence of decisions: what inputs were considered, what policies were evaluated, what exceptions were granted, who approved what, and why.
It’s a system of record for decisions, not just data. I think of it as a system of reasoning. (But I’m not promoting that as a phrase, because it’s easily confused with the reasoning that an LLM does).
Instead of systems only recording what decision was made, we should also record why it was made, in a structured way, every time.
- You have AI agents (or workflows) doing work
- Every time an agent makes or proposes a decision, the system logs:
- what it wanted to do
- what data it used
- who approved it
- why it was allowed or denied
That log is the decision trace.
This is about building a permanent, searchable memory of business decisions not just outcomes so humans and AI don’t keep relearning the same exceptions over and over.
As AI agents begin handling real workflows reviewing deals, resolving tickets, and more they run into the same gray areas humans face in everyday work.
Humans handle those situations using judgment and insider context built through experience, but agents don’t have access to that layer. They see the final state in the CRM, not the reasoning that led there.
Context graphs are supposed to solve this. By capturing decision traces as agents work, you build a queryable history of real-world precedents. Over time, exceptions become encoded knowledge. The organization stops relying on oral tradition and starts learning from its accumulated actions.
Smart folks like Jaya Gupta at Foundation Capital are making compelling cases. Startups building “systems of agents” could have a structural advantage because they sit in the execution path — they see the full context at decision time.