mirofish

Every organisation has a version of the same meeting. Someone proposes a change — new pricing, a policy, a public statement — and the room tries to imagine how people will react. It is guesswork performed by five people who share assumptions, and its main weakness is that nobody in the room thinks like the people who will be angry.

MiroFish attacks that with multi-agent simulation. You describe the scenario in plain language, attach supporting material, and several AI agents work through likely reactions from different positions, producing a structured prediction report.

This review covers what the five-stage workflow actually does, why the report's structure is the interesting design decision, how much confidence the output deserves, and where it beats simply asking a chatbot.

MiroFish, Described Plainly

• It simulates reactions to a decision — a price change, a policy, a campaign — using multiple AI agents rather than one answer.
• Files act as reality seeds. Attaching your own documents is what separates a grounded simulation from a well-written guess.
• Output is a structured report: executive summary, risk signals, narrative paths and follow-up questions.
• The follow-up questions may be the best part, because the value of a simulation is usually the objection you had not considered.
• Treat it as structured stress-testing, not prediction. Nothing here forecasts what will actually happen.

The Five Stages, And What Each Contributes

MiroFish describes a workflow of Seed Material, Knowledge Graph, Agent Simulation, Prediction Report and Deep Interaction. The names are neat; the value is unevenly distributed across them.

Seed Material And Knowledge Graph

Seed material is what you supply — the question, and any files that establish the actual situation. The platform's own guidance about using files as reality seeds is the right emphasis, because this is where a simulation is either grounded in your circumstances or floating free.

A question asked without documents produces a generic answer about how people generally react to price rises. The same question with your pricing page, your customer segments and last year's complaints attached produces something about your price rise. The knowledge graph stage structures that material into relationships the agents can reason over.

Agent Simulation

This is the part doing the differentiating work. Rather than one model producing one answer, multiple agents reason from different positions — which is a structural way of forcing the disagreement that a room of like-minded colleagues does not produce.

The honest caveat is that these agents share an underlying model, so their independence is real but bounded. They are not genuinely different minds; they are the same reasoning approached from different framings. That still surfaces more than a single pass, and it is less than the diversity of actual people.

Prediction Report And Deep Interaction

The report arrives structured as executive summary, risk signals, narrative paths and follow-up questions, with deep interaction available to push on any part of it.

Being able to interrogate the report matters more than the report itself. A static document invites acceptance; a document you can argue with becomes a working session.

Why The Report Structure Is The Real Design

The four sections are a considered answer to how simulation output goes wrong.

Risk signals force the analysis to name what could go badly rather than producing a balanced summary that offends nobody. Balanced summaries are the standard failure of AI analysis and they are useless in a decision.

Narrative paths is the sharpest idea here. It acknowledges that reactions are not a distribution of independent opinions but stories that spread — and a modestly held objection that turns into a shareable narrative does far more damage than a widely held indifference.

Follow-up questions is the section experienced users will come to value most. The point of stress-testing a decision is rarely the answer; it is discovering the question nobody in the room asked.

And the executive summary exists because these reports have to survive being forwarded to someone who will read four sentences of it.

The useful output is not 'people will react badly'. It is the specific objection you had not thought of, phrased the way someone angry would phrase it.

How Much Confidence The Output Deserves

Considerably less than the word 'prediction' invites, and this is the section that matters most.

Agent simulations produce plausible reactions, not probable ones. A language model reasoning about how a customer segment will respond is drawing on how such responses are generally written about, which is a real signal and not the same thing as evidence about your customers.

The failure mode is subtle and specific: the output is coherent, well structured and confidently expressed, and none of those properties correlate with accuracy. A report that reads like analysis gets treated like analysis.

Read it as a structured brainstorm with unusual discipline. It is genuinely better than a room of colleagues at surfacing angles nobody occupies, and genuinely worse than talking to five actual customers at telling you what will happen. Both statements are true simultaneously.

It also cannot know anything unusual about your situation that you did not put in the files. That is the whole argument for taking the seed material seriously.

• Good for: surfacing objections nobody in the room would raise
• Good for: pressure-testing an announcement before it is final
• Good for: generating the follow-up questions worth asking real people
• Weak for: anything presented as a forecast of what will occur
• Weak for: situations turning on facts you did not supply

Where To Actually Use It

The published use cases — campaign testing, pricing reactions, policy stress tests and market narrative — share a shape worth noticing. Each is a decision that is expensive to reverse, affects people who are not in the room, and normally gets made on intuition.

That is the sweet spot. A campaign can be tested cheaply with real audiences; a pricing change cannot be tested at all without making it, and a policy announcement gets exactly one launch.

The right sequence is to simulate first and then verify. Use MiroFish to generate the objections, then take the two or three most serious to actual customers or affected staff. That way the simulation is doing what it is good at — finding the question — while people supply the answer.

Using it the other way round, as the final check before shipping, is where organisations get into trouble.

1. Write the scenario as the real decision, including the parts that are awkward. A sanitised question produces a sanitised report.
2. Attach genuine documents — pricing, segments, prior complaints, the draft announcement itself.
3. Read the risk signals before the executive summary, so you engage with objections rather than the reassuring overview.
4. Push back through deep interaction on anything that seems too neat, and see whether it holds.
5. Extract the two or three sharpest follow-up questions and take those to real people.
6. Decide using the human answers, with the simulation as the thing that told you what to ask.
7. Keep the report afterwards and compare it against what actually happened — that is the only way to learn how much to trust the next one.

Three Terms Behind The Method

The approach is easier to judge once these are clear.

Multi-agent simulation

Several AI agents reasoning from different positions rather than one model producing one answer. It structurally forces disagreement, though the agents share an underlying model and are not genuinely independent minds.

Reality seed

Documents you attach that anchor the simulation in your actual situation. Without them you get a generic answer about how people react to things in general.

Narrative path

How a reaction spreads and mutates as a story, rather than as an isolated opinion. Often more consequential than how many people hold a view, since a shareable objection travels further than a common indifference.