Why We Built Prowir on Multi-Agent Systems Instead of One Very Good Prompt

2026-08-12 · The Prowir Team

Most pitches for multi-agent systems are arithmetic. One agent is clever, so five agents must be five times cleverer. That is not why we built Prowir as a set of specialized agents rather than one very good prompt. We built it that way because publishing in someone else's voice is not a single job. It is four jobs that actively disagree with each other, and a single prompt has no way to referee the disagreement.

Ask one model to research a topic, pick the angle, write it the way you'd write it, and then tell you honestly whether it worked. It will do all four in one pass. It will also quietly optimize for the one that is easiest to satisfy, which is fluency. You get something smooth, defensible, and slightly hollow — a piece that reads like it was produced rather than argued.

What multi-agent systems actually buy you

Not horsepower. The useful thing multi-agent systems give you is the ability to hold two incompatible objectives at the same time.

A drafting agent's job is to produce. A critic's job is to reject. Those are opposed incentives, and asking the same context to serve both is like asking an architect to stamp their own structural drawings. They'll approve them. Not because they're dishonest, but because the person who made the choices is the worst possible person to notice the choices they didn't consider. Separation isn't ceremony. It is how you get an actual "no."

The same logic shows up everywhere real work gets checked. A clinical trial has an endpoint committee that didn't run the study. A manufacturer has QA that doesn't report to the line manager. A consulting firm has a partner review that exists specifically to embarrass the deck before the client does. In each case, the reviewer is deliberately given less context and a different mandate — and that is the point, not a bug.

The failure mode the demos skip

Here is the honest part, because we said we'd be honest about where AI falls down: most multi-agent systems fail for boring organizational reasons, not because the model isn't smart enough.

Berkeley researchers studied this directly in Why Do Multi-Agent LLM Systems Fail? — a paper that catalogued failures across real multi-agent traces and grouped them into a taxonomy of fourteen distinct modes in three families: badly specified roles, agents drifting out of alignment with each other, and verification that rubber-stamps instead of verifying. Almost none of it was raw reasoning failure. It was agents losing the thread on handoff, restating the task slightly wrong, or a "reviewer" agreeing with whatever it was handed.

Cost is the second honest caveat. When Anthropic published its own multi-agent research system, the multi-agent setup meaningfully outperformed a single agent on research tasks — while burning roughly fifteen times the tokens of a normal chat exchange. More agents is not free, and it is not automatically better. It is a trade you should be able to justify per task.

So the design question is never "how many agents." It's: which disagreements are worth the coordination cost, and what exactly gets passed between them.

How we split the work

We split Prowir along the seams where the judgment actually differs.

Voice is its own agent, and it doesn't write. It maintains a model of how you specifically build an argument — sentence rhythm, what you hedge, what you refuse to say — learned from things you've already written. It has no opinion about the topic. That isolation is deliberate; the moment a voice model starts contributing content, it starts flattening toward the average of everyone's content.

Research doesn't draft. It gathers claims with sources attached and hands over an artifact, not a vibe. A draft agent that does its own research will invent the convenient statistic every time, because inventing is cheaper than finding.

The critic gets a narrow mandate and permission to kill. Its question is not "is this good writing." It's "would a skeptical senior person in this field respect this, or skim past it?" A product leader, a structural engineer, an oncologist — each has a bar for what counts as a real point, and generic competence sits well below it.

Handoffs are explicit. Every transfer between agents is a structured artifact rather than a conversational summary, because summary-by-conversation is precisely where the Berkeley failure modes live.

Why this matters now

The cost of producing a plausible post has gone to roughly zero, and everyone's feed shows it. Volume is no longer evidence of anything. What is scarce — and getting more scarce — is a piece of writing that a peer reads and thinks, that person has actually done this work.

A single prompt is very good at producing the first thing and structurally incapable of producing the second, because it has no mechanism for self-doubt. Multi-agent systems are how you build the doubt in on purpose: one part that produces, one part that refuses, one part that remembers you don't talk like that.

None of this is magic, and we won't pretend a system can have a point of view. Yours has to come from you. The architecture just makes sure it survives the trip from your head to the page.

If you'd rather publish the thing you'd actually say than the thing a model finds easiest to say, Prowir's beta is open.