Deep Research
Six agents searching at once, and one report that cites what they found
A standard analysis sends one search. Deep Research plans the questions first, runs six web-search agents concurrently, and synthesises what came back into a single report with its sources attached and a confidence score on top.
Why a planner comes first
The failure mode of automated research is not bad searching, it is searching for the wrong thing well. A single query against a vague idea returns whatever the index thinks the words mean, and the report that follows is a summary of a misunderstanding.
Planning splits the question before any search runs: who the buyer is, who already serves them, what the regulation looks like, what the substitutes are. Six narrow questions return usable evidence where one broad one returns noise.
Why the agents run in parallel
They are independent. Nothing the competitor search finds changes what the regulatory search should ask, so running them in sequence buys nothing and costs the sum of their latencies.
The concurrency is bounded rather than unlimited, because the ceiling that remains is the provider’s rate limit, and firing everything at once trades a wait for a failure.
What synthesis actually does
It reconciles rather than concatenates. Six agents routinely return overlapping claims, and two of them will contradict each other on something specific — a market size, a competitor’s pricing, whether a rule is in force. Concatenation hands that contradiction to you unmarked.
The synthesis step keeps the disagreement visible and attributes each claim to where it came from. Every figure in the report traces to a source you can open, which is the difference between research and assertion.
The confidence score, and when to distrust the report
The report carries a confidence score, and its purpose is to be low sometimes. A pipeline that always reports high confidence is telling you about its own design rather than about your idea.
Confidence falls when the sources thin out, when they disagree, or when the searches returned commentary rather than primary material. A low score is not a failure of the run — it is the run telling you that the evidence for this particular question is not there, which is itself worth knowing before you build.
Available on the recurring plans.