Skip to main content
A research agent usually needs two distinct capabilities:
  1. Retrieval of current or authoritative sources.
  2. Model reasoning to compare, summarize, and explain the evidence.
Do not use model inference alone for claims that depend on current information.

Choose the research surface

Use a Skill when the workflow is reusable and the agent should follow an established process. Use direct APIs when the application needs to control source selection, sequencing, caching, or ranking itself.

1. Define the research question

Specify:
  • The exact decision or claim the research should support.
  • The required time range.
  • Preferred source types or domains.
  • Geographic or language constraints.
  • The expected output and citation format.

2. Retrieve before synthesizing

Use search results for discovery, then fetch the source pages needed as evidence.
Do not treat snippets, model memory, or a generated answer as equivalent to a primary source.

3. Compare sources

For material claims:
  • Prefer official documentation, filings, original announcements, or primary datasets.
  • Record publication dates and source URLs.
  • Identify contradictions instead of silently merging them.
  • Mark claims that remain uncertain or depend on one source.

4. Use a model for synthesis

After retrieval, pass the relevant evidence to a supported model. Ask it to separate:
  • Verified facts.
  • Interpretation.
  • Open questions.
  • Recommendations.
The model should not invent citations or claim access to sources that were not retrieved.

5. Verify the output

Before delivering the result:
  • Confirm each important citation resolves to the intended source.
  • Recheck time-sensitive numbers.
  • Remove unsupported claims.
  • State the research cutoff time.

Minimal documentation set

An agent building this workflow should normally read only: Do not ingest the entire API reference unless the task truly needs broad endpoint discovery.

When AIsa may not be necessary

A single specialized search provider may be simpler when the workflow uses one source type, does not need reusable Skills, and already has an established model and billing integration. For product-fit guidance, see When to use AIsa.