Ask an AI one question and you get one answer, shaped by one point of view: yours. That feels efficient until you notice what is missing. A single query carries your own blind spots straight into the research, and blind spots are invisible to the person who has them. If you don't already know a field well, you won't even know which questions you failed to ask.
There is a sturdier shape for research, and it doesn't depend on any one tool. You gather several expert points of view instead of one, get them to disagree on purpose, map exactly where they clash, then check every claim against its source before you believe a word. I'll walk through each step, and show where you'd use quiet worker agents versus a team that argues out loud.
A panel of perspectives, not one prompt
Instead of firing off a single prompt, you spin up a small panel and give each member a distinct expert role. A common set is five: a practitioner who cares how a thing behaves in daily use, an academic who cares what the evidence actually shows, a skeptic whose whole job is to poke holes, an economist who follows the money and the running cost, and a historian who remembers how similar ideas played out before.
Each lens looks for different things, so each one catches a gap the others walk straight past. Ask "are voice AI agents worth building a project around?" as one prompt and you get a tidy, confident summary. Ask the same question through five lenses and the picture fractures in a useful way. The practitioner reports what breaks once you leave the demo. The economist flags where the running cost quietly balloons. The skeptic names the parts that are hype. The historian points out that we have seen this wave before under an older name. The academic separates the flashy demos from what the evidence supports. The single prompt told you none of that, not because the AI is weak, but because you only asked from one chair.
I didn't invent this. The multi-perspective approach is a documented research method — one well-known version is STORM, from Stanford — and the idea underneath it is old and reliable: more independent viewpoints produce more complete work. What is new is that you can now summon those viewpoints on demand instead of hunting down five human experts.
Where they disagree: the contradiction map
Getting five answers is not the point. Five answers sitting in five folders is just more reading. The value appears when you force the lenses to read each other.
The second move, then, is a contradiction map. You hand every perspective's output back to the group and ask a blunt question: where do you contradict each other, and whose evidence is actually strong? The academic might claim adoption is climbing fast. The practitioner might counter that the tools fall over the moment they leave a controlled demo. Rather than quietly averaging those into a mushy "it depends," you write the disagreement down as a disagreement, and you note which side brought evidence and which side brought a hunch.
That map is often the most useful thing the whole exercise produces. It shows you where the topic is genuinely settled and where it is still a fight, which is exactly the information a confident one-paragraph summary hides. When two lenses agree and three object, you have learned something you would never get from a source that only ever agrees with itself.
Verify before you believe
Any AI research has one uncomfortable property, whether it came from a single prompt or from five lenses: the first pass will contain claims that are simply wrong. A number that got garbled in transit. A source that does not say what it was cited as saying. A confident sentence with nothing underneath it.
So you don't ship the first pass. You run a verification round where each claim is checked against its primary source and sorted into one of three bins. Confirmed: the source backs it. Corrected: the source says something close but different, so you fix the wording. Demoted: no solid source stands behind it, so it drops in status or gets cut. A claim that survives this and is backed by several lenses earns high confidence. A claim that only one lens asserted, with no source to confirm it, gets flagged as shaky instead of being quietly presented as fact.
The finished report then carries its own trust labels. Each key finding is ranked by how reliable it is and shows which lenses backed it and which pushed back, so you are never asked to take the whole thing on faith. You can see at a glance the difference between "this one is solid and several lenses agree" and "this one is thin, handle with care." This is the same instinct behind a rule I hold to everywhere else: don't ask if it's done, make the agent prove it works.
One more habit is worth stealing. A good verified report also names the assumption the whole thing rests on, and the lens it forgot. In the voice-agents example, all five lenses looked at the question from the owner's chair: adoption, productivity, return on the investment. None of them sat in the seat of the customer or the frontline employee who actually has to use the thing. Naming that missing lens is itself a finding. You add the sixth chair and run another version.
Sub-agents or an agent team
There is a design choice underneath the panel idea, and it is worth understanding because it changes both the power and the price.
When you run the panel as sub-agents, you have one main session — the assistant you are actually talking to — and it dispatches several workers. Each worker goes off, does its research, and reports back. The limit that matters: the workers cannot talk to each other. The main session gathers the separate reports and stitches them together. That is cheap, quick, and exactly right for a research panel, where you want independent takes precisely because they never influenced one another. For when to reach for them, I've written separately about sub-agents and where they fit.
An agent team is a different animal. There, the members can talk to each other, not only to the main session. You can point them at a decision and have them debate, arguing back and forth until they reach some kind of consensus, instead of each filing a report in isolation. That is more powerful when you need a judgment call rather than a survey, and it costs more, because all that back-and-forth burns more compute. Agent teams that debate to a consensus earn their keep when a lone answer would be too fragile to trust.
The rule of thumb stays simple. Independent research, where you want the blind-spot coverage of lenses that never colluded: use sub-agents. A contested decision, where you want the perspectives to stress-test each other in real time: use a team that debates. Some assistants also ship a built-in deep-research mode that can spin up hundreds of worker agents at once for a single question. That is the sub-agent pattern at scale — see dynamic workflows — and it returns a broad sweep, but sheer breadth is not the same as the disagree-then-verify discipline above.
The lesson is the principle, not the tool
Strip away the specific method and the specific software and one idea remains, the one actually worth keeping: the more independent perspectives you can get to disagree with each other, the more complete and more trustworthy your research becomes. The tool is incidental. The discipline is the whole thing.
That points at a genuinely useful move for anyone working outside their own expertise. If you don't have the domain knowledge, borrow it. Stand up a handful of role-played experts whose only job is to find the gaps you cannot see, get them to argue, then make each one prove its claims against a real source. What you get back is close to what a room full of specialists would have told you, and you can see exactly where they failed to agree.
Next time a question matters enough that one confident answer would be dangerous, don't ask once. Ask from several chairs, write down where they clash, and believe only what survives the check.