Agents that answer with sources
A claim asks for trust; a cited fact earns it. Which tools return references, how to ask for citations, and how to design answers people can check.
2 min read
An agent’s answer is only as useful as it is checkable. “The customer agreed to the new terms” is a claim. “The customer agreed to the new terms in their reply of March 3” with a link is a fact your user can act on. Here is how to build agents that cite.
Why citations change behavior
- Users trust what they can verify. One click to the source turns doubt into confidence.
- Models read before they write. An instruction to cite every claim pushes the model to ground each one in a tool result.
- Mistakes become visible. A wrong answer with a source is easy to catch; a wrong answer without one slips through.
Start with sources that carry references
Citations need something to point to. Choose tools that return it:
- Web search returns the URL of every result.
- Perplexity returns a synthesized answer with its citations.
- iGPT returns email context with citations to the source messages and attachments.
- Workflows can include identifiers and links in their output, so each item points back to its record.
Ask for citations explicitly
Conceptual example.
instructions: `
Answer using only information returned by your tools.
After every claim, cite its source in brackets with a link.
If the tools did not return enough to answer, say so.
`The last line matters as much as the first two: permission to say “I could not find that” removes the pressure to invent.
Design the answer for checking
- One claim, one citation. A paragraph with a single link at the end hides which sentence it supports.
- Quote when precision matters. For commitments, prices and dates, a short quote beats a paraphrase.
- Separate fact from inference. “The thread shows X” is different from “this suggests Y”. Ask the model to label which is which.
- Render links properly. In your interface, make citations clickable and show where they lead.
Test for it
Add questions to your test set whose answers are not in any source. A good agent says so. An agent that answers anyway needs stricter instructions, or a different model for that task.
An answer without a source asks for trust. An answer with one earns it.
See cited answers in action on the iGPT page, or read how to build a research agent.
