Research Agent
Ask a capable colleague what is currently known about a narrow question and they do not answer from memory. They search, skim, follow citations, discard the thin sources, and come back with a short written account and a list of what it rests on. A research agent is software built to run that same errand.
How the loop spends its time
It wraps a language model in a loop with search and retrieval tools, and spends most of its run gathering rather than generating. A cycle looks like: break the question into sub-questions, issue searches, fetch and read documents, record what each one actually supports, notice a gap, search again. The loop ends when the sub-questions are answered or a step or token budget runs out. State machinery such as LangGraph earns its place here, because the agent has to remember which sources it already rejected and why.
Why it beats a single call
One model call can only report what it absorbed during training, confidently and without a trace. The value of a research agent is the chain of evidence it leaves behind — each claim tied to a document it retrieved — which makes the output checkable rather than merely fluent.
Where it goes wrong
The failures are specific:
- The opening plan fixes a framing that later evidence contradicts, and nothing in the loop revisits it.
- Source notes fall out of the context window, so the write-up cites material the agent no longer holds.
- Retrieval returns plausible but weak pages, and the synthesis inherits their errors with the same confident voice.
A research agent is an agent specialised for gathering and synthesising information: it breaks a question into sub-questions, searches and retrieves documents, reads them, and assembles a cited written answer. Most of its run is spent collecting rather than generating. It matters because it replaces a single model call's unsourced recollection with a traceable chain of evidence the reader can check.
Picture asking a thorough assistant to find out what is known about something. They don't answer off the top of their head. They look things up, read what they find, throw out the thin material, and come back with a short written summary and a list of where it came from. A research agent is software that runs that errand.
Given one question, it searches, reads, notices what it still doesn't know, and searches again, until it can write up an answer with its sources attached. The sources are the point: you can go and check the claims yourself instead of taking the answer on trust.