The Research Pipeline

A single web search produces a single answer of unknown quality. The pipeline exists because trustworthy research is a process: gather widely, then attack what you gathered, then fill the holes the attack exposed, then write up only what survived. The platform runs that process unattended and gives you the artifact at the end.

Queueing a Task

Research tasks queue exactly like coding tasks: a file in work/research/queue/, written by the master agent when you ask a research question, by Impulse when a goal needs information, or by you directly. A good task reads like a brief for an analyst: the question, why it matters, and what a useful answer would settle. The pipeline node ticks on its own schedule and works the queue one task at a time, each task running as long as the question deserves.

The Passes

Search casts wide: the session works the web for relevant sources on the question, gathering claims and evidence rather than settling for the first plausible answer. Skeptical review is the pass that earns the word verified: the same session turns adversarial, checks claims against each other, hunts for contradictions, and asks what a doubter would ask of every finding. Improve repairs what review exposed, searching specifically for what is missing or contested. Synthesize writes the final document from what survived.

The continuous session ties the passes together: the skeptic remembers exactly where each claim came from, and the synthesis pass knows which findings stood up to challenge and which were repaired. The end product reflects the whole journey, conclusions carry confidence levels, and claims carry their sources, so you can see at a glance what is solid and what is best-available.

Where Results Go

Every task produces a verified knowledge document you can read. With kb_insert set in the pipeline settings and a target kb_name, the findings also flow into a knowledge base, paragraph-chunked, embedded, and immediately searchable by every agent in the system. That one setting turns the pipeline into a knowledge factory: research your support team's recurring topic and the answers become support material the same day; research your market and the social media preset draws on verified findings for its post proposals, which is exactly where its paced publishing gets its substance.

Recurring Research

The pipeline gets more valuable when it stops being a one-off. Put a standing research goal in the Goals tab, keep an eye on competitor pricing, watch for developments in a topic, and Impulse queues the work on its own initiative, so your knowledge bases stay current without you remembering to ask. The combination of a goal, the pipeline, and kb_insert is a self-refreshing intelligence function: questions your business always needs answered, re-asked and re-verified on a rhythm, with the answers always one query away for every agent and for you.

Reading the output efficiently is its own small skill: start with the conclusions and their confidence levels, follow the sources only where a claim surprises you, and treat anything marked uncertain as the seed of the next task. The documents are built to be skimmed by a busy owner and trusted by the agents who cite them.

Model and Schedule

The pipeline has its own card in the Config tab's Models panel for model and tick. Research rewards reading comprehension and skepticism, so mid-to-strong models do it justice, and the tick simply sets how often the queue is checked. For interactive work, the one-shot research tool runs the same multi-pass discipline from the UI while you watch.

Working With It Well

Ask real questions rather than keywords, "what do competitors charge for X and how do they justify it" gets a better document than "competitor pricing". Set kb_insert when the topic will come up again, one-off curiosities can stay as documents, recurring themes belong in a base. And chain it: the best research briefs often come out of conversation with the master agent, where you can sharpen the question before it hits the queue, and the best follow-ups come from reading the confidence levels in the result, anything marked uncertain is a ready-made next task.