How a network of AI agents produces an analysis that a single model can't — explained from the ground up.
Ask one capable AI model a hard, contested question and it gives you a single answer. It surveys what it knows, weighs it, and hands back one confident synthesis. That sounds like what you want, but the synthesis hides the most useful information: where informed people actually disagree, which factor really decides the outcome, and how the same fact means different things to different stakeholders. A single voice has to flatten all of that into one position. The flattening is the loss.
The mesh is built to avoid the flattening. Instead of one analyst, it runs a population of specialist agents — think of them as a CFO, a CISO, a procurement lead, a systems-integrator partner, a model-risk officer, a frontline operator — each reading the same research through the lens of a different expertise. They write up what they see, then react to each other: agreeing, pushing back, qualifying, reframing. The output isn't a consensus paragraph. It's a structured map of how the question looks from many expert angles at once.
Three properties describe how that works. They're worth taking one at a time — but the punchline, which the rest of this builds toward, is that two of them are really the same thing at different scales, and the third is the one that makes the mesh genuinely different from a single model.
Give a generic summarizer a pile of research and it produces a generic summary — the obvious points, evenly weighted. Give the same research to a specialist and they immediately notice things the summarizer would skip, because their expertise tells them what's load-bearing.
That's the first property: attention is differentiated noticing. Each agent has, in effect, a different sense of what's relevant. In a recent run on enterprise AI adoption, this was vivid. The finance and procurement agents zeroed in on contracting terms, cost attribution, and renewal quality. The security and model-risk agents zeroed in on audit logs, permissions, and the need to re-validate when anything changes. The operations agents zeroed in on who handles the exceptions when the system hits a case it can't manage. Same documents — but each specialist pulled different things into focus, the way different experts reading the same contract circle different clauses.
(One honest note: these specialists are the same underlying model steered into different roles, not separate AI systems. The different focus is real and useful, but it comes from role and perspective, not from architecturally distinct minds. Worth saying plainly, because it's the first thing a skeptic asks.)
Here's where the mesh stops resembling anything a single model does.
Once each specialist has weighed in, they engage each other — and the system preserves that engagement instead of dissolving it into an average. One agent says a point is the key mechanism; another replies that it's only true in regulated industries; a third says it looks like a strong recurring-revenue signal but only if you separate out the one-time setup work; a fourth says governance matters less than who's contractually on the hook when something breaks.
That's polyphony: structured disagreement, qualification, and reframing among expert voices, held in place rather than collapsed. It's not just "many opinions." It's a record of where perspectives reinforce each other, where they contradict, and — most valuably — where the same claim means different things depending on who's looking at it. "This workflow is ready to scale" means something different to a CFO than to a security officer than to the team that runs it day to day. A single-voice answer can only pick one of those meanings. The mesh keeps all of them, and shows how they relate.
This is the property that can't be faked by a clever prompt. A single model can be told to "weigh what matters" or "rank the important points" — those are doable alone. It cannot produce genuine, independent disagreement among separately-reasoning viewpoints, because by construction it is one viewpoint. Polyphony is the thing you can only get from a network.
The third property is about importance — but importance that emerges rather than being declared. A point doesn't become central because one agent insists on it. It becomes central because other, unrelated specialists independently pick it up, build on it, or keep returning to it in their own terms.
That's salience: relevance revealed by propagation. The signal isn't volume; it's reach across unlike perspectives. Which claims travel from the finance lens to the security lens to the operations lens? Which questions keep resurfacing no matter where you start? In the enterprise-AI run, the point that traveled furthest — surfacing in nearly every specialist's reasoning regardless of their angle — was that whether these systems succeed depends far less on how good the AI model is and far more on the unglamorous plumbing around it: governance, integration, ownership, contracts. It became the center of gravity not because anyone asserted it loudest, but because every independent lens kept arriving there.
Notice that attention and salience are doing the same kind of work. Attention is one specialist deciding what's relevant; salience is the whole network deciding what's relevant. They're the same operation — relevance-weighting — at two different scales: inside one agent, and across all of them. A point that's salient is just a point that many agents' attention converged on.
Polyphony is the genuinely different one. Attention and salience are about weighting things; polyphony is about preserving difference — refusing to let the weighting erase the contradictions and the context-dependence. That's why, of the three, polyphony is the one a single model fundamentally cannot reproduce, and the one that makes the mesh worth building.
So, compactly:
Or in one line: attention is what a specialist notices, salience is what the network amplifies, and polyphony is the difference the system refuses to erase.
The mesh produces structure — what the specialists notice, where they disagree, which cruxes the whole network keeps circling. That tells you what the real possibilities are and where the genuine uncertainty lives. But it stops short of a probability.
That's the second stage, called Many-Worlds. It takes the structure the mesh produced — the key dimensions, the live disagreements — and runs a simulation across them, sampling the uncertainty thousands of times to produce a distribution of outcomes. The mesh decides what the contest is and where it lives; Many-Worlds prices it.
One line captures the whole system: the mesh refuses to collapse the disagreement; Many-Worlds measures how it's likely to resolve.