2026 · 3 min read
What Is MiroFish? Synthetic Audiences, Explained
What a synthetic audience actually is, and how a simulated population becomes decision-grade research.
Most market research asks real people what they think. You recruit a panel, you run a survey or a focus group, you wait, and you hope the sample resembles the world. It is slow, it is expensive, and the people you reach are rarely the people you most need to understand.
MiroFish takes a different path. Instead of sampling a population, it generates one — a synthetic audience of named, individuated agents, each with a history, a disposition, and a way of speaking — and then lets that population react to whatever you put in front of it. The output is not a number on a dashboard. It is a chorus of distinct voices responding the way a real audience would, captured before anything reaches a real human at all.
From source material to a living population
Every MiroFish simulation begins with what we call a reality seed: the curated source material that defines the world the agents live in. This might be a clinical care pathway, a product concept, a trading platform, a film treatment, or a body of real social conversation. The seed is the ground truth — the facts, constraints, and texture the synthetic population takes as given.
From that seed, MiroFish generates a synthetic ecosystem of agents. These are not interchangeable bots. Each one is a constructed persona with its own profile, drawn to populate the specific cohort the question requires — a community of patients managing a chronic condition, a population of funded-account traders, an audience for an independent documentary. The population is designed, not scraped, which means it can model people who would never answer a survey.
Running the population across conditions
Once the ecosystem exists, the agents are run across parallel interaction conditions. The same population can encounter a message in different forms — static, turn-based, real-time — and the differences in how they respond become the finding. Because the whole population is synthetic, you can run conditions side by side that no real-world study could afford: every agent sees every variant, with no recruiting, no scheduling, and no waiting.
The agents interact, react, and generate language. They express confusion, enthusiasm, skepticism, and misunderstanding. They surface the objection nobody on the team anticipated. They reveal where a care plan loses a patient, where a pitch loses a trader, where a story loses an audience — early, cheaply, and in their own words.
From simulation output to decision-grade findings
A population of talking agents is raw material, not an answer. The final and most important step is editorial: translating thousands of synthetic responses into a clear, executive-grade read of what happened and what it means. This is where two decades of intelligence work meet the simulation — the discipline of turning a flood of signal into a small number of recommendations a decision-maker can actually act on.
The technical infrastructure underneath — multi-agent orchestration over an LLM backend, model selection and routing balanced against cost and privacy constraints, a persistent memory layer so agents retain who they are across an interaction — exists to serve that final translation. The machinery is in service of meaning.
Why synthetic, and why now
Three things make this possible now that were not possible a few years ago: language models fluent enough to sustain a believable individuated persona, orchestration tools that can run many such personas in parallel, and memory systems that let an agent stay coherent across a long interaction. Together they make it feasible to stand up an entire audience in software and listen to it.
The point is not to replace real people. Real cohorts still matter — in regulated domains like health, synthetic simulation is best run ahead of a real cohort, not instead of one. The point is to fail early and cheaply in simulation, so that what reaches real people is already better. A synthetic audience lets you ask the awkward question, test the risky message, and watch the misunderstanding unfold before it costs anything in the real world.
There is, finally, a quieter thread running underneath all of this. Building convincing synthetic people forces you to think hard about what makes a person legible — their narratives, their contradictions, the meanings they carry. That is older territory than AI, and it is where this work draws as much from depth psychology as from engineering. MiroFish is a technical system, but the questions it raises are human ones.