How research systems work

Why several agents instead of one?

Ask one AI to read a stack of interviews and hand back the findings, and you get a fluent summary with no way to check it. Everything happened in one step, out of sight.

Splitting the work fixes that. One agent finds passages where a participant describes their own thinking. Another drafts a short summary of each. Another looks for what is missing. Each one has a narrow job and sees only the material that job requires, so when something goes wrong you can tell which step produced it.

Agent

An AI given one specific job and a written set of instructions for doing it. Not a separate program you buy — a role you define.

What does deterministic mean here?

Ask an AI the same question twice and you may get two different answers. That is not a defect. It is how these models work, and it is part of why they are good at noticing patterns in messy human language.

But it is a terrible property for anything that needs to be reliable. So the AI never controls the process itself. Ordinary software — the predictable kind, which does the same thing every time — decides what runs in what order, what each agent may read or write, and what gets recorded. The AI interprets. The software keeps order.

Deterministic

Same input, same output, every time. Ordinary software is deterministic. AI models are not.

Probabilistic

The opposite: results vary between runs. Useful for interpretation, unsuitable for control.

What is a trace?

Every claim the system produces carries a record of where it came from: which interview, which passage, which agent proposed it, under which version of the instructions, and what a reviewer decided about it.

This is what makes a finding checkable. Someone reading a conclusion months later can follow it back to the exact words a participant said, rather than taking it on faith. Without it, a plausible-sounding sentence and a well-supported one look identical on the page.

Trace

The recorded path from a finished claim back to the original evidence.

Provenance

Where a piece of information came from and what has happened to it since.

What is a review gate?

A point where the work stops until a person looks at it. Nothing moves past a gate on the system's own authority.

The agents propose; the researcher disposes. A proposal can be accepted, edited, sent back for a source check, or rejected outright. The decision is recorded alongside the work, so later readers can see not only what was concluded but who agreed to it.

Review gate

A required human approval between one stage of the work and the next.

What breaks without this?

Three things, in rough order of how quickly they arrive.

Findings that cannot be defended. A conclusion arrives with no way to check it. It gets copied into a deck, then a strategy document, and somewhere along the way it stops being a guess and starts being a fact — without anyone deciding it should.

Cost that grows without limit. An AI asked to search an unorganized pile of documents has to read all of them to find anything. That is billed by the word. Structure is not tidiness here; it is the difference between a system an organization can afford and one it switches off.

Judgment quietly handed over. A well-formatted summary reads like analysis whether or not it is. The risk is not that AI replaces researchers. It is that people stop checking, because the output looks finished.