When people talk about “context,” they usually mean “some text the model should see.”
That’s a starting point, but it’s not a useful definition for agent systems.
Agents don’t just read text. They make moves. And when you make moves, you need more than words.
Here’s a practical definition:
Context for AI agents is text plus the fields that make it safe to use.
What those fields are
In practice, the fields you end up needing are predictable:
- Source: where did this come from?
- Time: when was it published, and what window does it cover?
- Evidence: is this a quote, an observation, a summary, or a derived claim?
- Confidence: how strong is the grounding?
- Identity: can we refer to the same underlying thing tomorrow (stable IDs)?
- Routing: why was this included for this agent and this job (streams)?
If you have those, “context” becomes something an agent can reason over and a team can audit.
If you don’t, the model has to infer them from prose, which means it will sometimes invent them.
A small example
Imagine an agent answering a simple question:
“What’s our refund policy?”
If you give it the policy text, it can quote it.
But if you want it to operate, you need more:
- which version is in effect,
- when it was last updated,
- whether there are regional variants,
- whether a customer’s case falls under an exception.
That’s context. Not just text.
This is why Hanging Context focuses on public panels and JSON twins: it’s the public, citeable shape of a context system. And it’s why Synorb exists as the retrieval layer: it delivers the structured objects that production agents actually need.