Main idea

Many deep research systems follow a "search → scrape → summarise → write" architecture. Even when they run several searches, these remain branches of a plan set very early on. Research there is a phase that comes before generation.

A genuinely adaptive loop requires the agent to maintain a state: the scope, the stabilised knowledge, its provenance and level of support, the contradictions, the uncertainties, the gaps identified, the searches already run, the strategies that yielded nothing, the areas judged saturated. That state is what lets it answer the only useful question at each step: which search now maximises the chance of a gain in knowledge?

Depth is no substitute for that state. An agent can read a hundred documents while repeating the same biases, explore enormously without ever looking for the source that contradicts its central conclusion, accumulate a gigantic context without knowing what remains uncertain.

Why it matters

It shifts how you evaluate an AI research tool: not how many sources it reads, but what it knows about what it has understood and what it is missing.

And it says what has to be built on top of a research pipeline: a memory of the state of knowledge, updated at each iteration and reread before the next.

Nuances and limits

Keeping that state costs, in tokens as in design, and on a simple question a linear search is enough.

And a badly kept state — uncertainties not updated, areas wrongly declared saturated — leads the agent to choices that are more confident but no more correct.

Open questions

• In what form does this state stay readable for a human who wants to check the agent's choices?