Context Graphs for AI Agents

An AI agent plans and carries out multi-step tasks: it updates records, sends messages and opens tickets. Each step depends on facts about the company. A context graph gives the agent those facts as connected nodes with sources, so it can check before it acts.

This article is about how agents use a graph. For what a context graph is and how it differs from a knowledge graph, read What Is a Context Graph?

• Locate: find the entities in the task, such as the customer, the project or the person who asked.
• Traverse: follow edges to related facts, such as open tickets, the account owner and recent decisions.
• Check: read the source and date on each fact, and drop facts that are out of date.
• Act: take the next step, citing the facts it used.
• Write back: record what it did, so later agents and people see it.

Vector search finds text that looks like the question. That works for a single answer from a single document. Agents often need multi-hop facts: the owner of the account behind this ticket, or the decision that replaced last quarter's plan. A graph answers these by following links, and it returns facts with their sources, so the agent's output can be traced.

An agent should see only what the person it works for may see. A context graph that keeps each source's permissions enforces this on every query. Sources and dates on each fact also let a reviewer check why the agent acted. This is a direct way of reducing AI hallucinations.

Want to give your agents a context graph? See the East River Labs product, or talk to us.