Skip to main content
Users don’t trust an agent because it sounds confident. They trust it when they can see why it did what it did — the specific piece of context behind the action. Provenance is that trail. Without it, every answer is “trust me,” and adoption stalls exactly where the stakes are highest: legal, healthcare, finance, infrastructure.

​
The problem provenance solves

As you collate more documents, finding which reference the agent used gets harder:
  • The agent read across several documents; which sentence actually drove the action?
  • Was that fact current, or an old copy from a stale file?
  • Who asserted it, and how confident should we be?
When you can’t answer those on demand, you can’t put a proof-of-work in front of the user — and without proof-of-work, they can’t trust the output. That single gap caps how much you can hand off to the agent.

​
Attribution belongs on the fact, not the document

The fix is to attach provenance to each fact, so it travels with the fact into every answer:
1

Address every fact

Each entity has a stable id. An answer can cite service:checkout precisely — not “somewhere in the payments folder.” → entities
2

Record who asserted it, and how confidently

Attach the source and confidence when a fact is written — a connector, an import, an agent, a human. In Open Index, writes carry a provenance block (asserted_by, confidence), and bulk imports attribute a whole batch at once (--asserted-by import:jira). → populating entities
3

Make retrieval citable

Because the agent retrieves discrete, addressed records — not a blob of merged text — every retrieved fact can be shown to the user with its id and source. The proof-of-work assembles itself from what was actually fetched. → Context search accuracy

​
Provenance is also an audit tool for you

The same trail that builds user trust lets you debug the agent. Local fetch analytics record what context each client retrieved, how often, which queries returned nothing, and where latency or failures cluster. When an agent acts wrong, you can see the context it pulled — instead of reconstructing what it “must have” used. → MCP analytics
Zero-result searches are a gift: they show precisely what your users asked for that your knowledge base couldn’t answer — i.e. what to model next.

​
Checklist

  • Every fact is individually addressable and citable.
  • Writes carry source + confidence (who asserted this, how sure).
  • The agent can produce a proof-of-work listing the exact records it used.
  • You can audit, after the fact, what context an action was based on.

Next: Human steering

Reducing the autonomy tax now that context is trustworthy.