Running research as parallel agent lanes

by Dev Ashish · 9 October 2026 · 2 min read

Split a research question into independent lanes, give each AI agent one lane, the facts you already know, a rule that every claim cites a source, and a fixed shape for its answer. Run the lanes at the same time, then merge the answers and check a sample of the citations before trusting any of it.

When we research something at Codeclowns, a market, a competitor, a technical choice, we rarely ask one AI agent the whole question. We split it into lanes and run several agents at once.

What a lane looks like

A lane is one narrow question with edges. For a pricing question the lanes might be what competitors charge, what buyers pay today for the alternative, what our costs are per unit, and what the buyer's budget owner can approve without a meeting. Each lane goes to its own agent.

Every agent gets the same seven things in its brief:

1. Its role and one objective. One agent, one job.
2. Its lane, and what its siblings cover, so nothing is researched twice.
3. Where to start: the files, pages or data to open first.
4. What we already know, stated as facts, so it doesn't re-derive them.
5. Where to look for the rest.
6. The exact shape of its answer, so the answers merge mechanically.
7. When to stop, and what not to report again.

The fifth and sixth matter more than they look. An agent that is told what we already know spends its effort on what we don't. An agent that must answer in a fixed table can be compared line by line with the others.

Why parallel beats one big prompt

A single agent asked a broad question spreads itself thin. It finds a plausible answer to the first part and treats the rest as solved. Separate lanes each go deep, and when two lanes disagree, the fixed format makes the disagreement obvious instead of buried in prose.

It is also faster in plain clock time: five lanes take about as long as the slowest one.

The rule that keeps it honest

Every claim cites a source, with a date. Market facts expire, and a number without a date quietly turns into a claim about today. Before any finding feeds a decision, a person checks a sample of the citations. When a citation doesn't hold up, the whole lane is treated as suspect, not just that line.

Where the answers go

The merged findings are written into our product hub, dated, with their sources. The next time anyone, person or agent, asks a related question, they start from that file instead of from zero. That is most of the value: research that compounds instead of being redone.

This is the second step of how we work. The third, where two agents argue opposite sides before we decide, is how we find out whether the research actually supports what we want to do.

Questions

← All notes