Is MiroFish any good?

MiroFish is good at one thing: turning a document into a simulated public that argues about it, so you can see objections, factions and reactions you had not thought of. It is not proven at forecasting: the project publishes no accuracy figures, results can change between runs and models, and an open bug report shows its report agent can invent quotations. Used as a scenario lab, with the evidence checked, it is useful; used as a prediction, it is not.

Written by MiroFish.co.uk, an independent hosted service built on the open-source MiroFish project, so we have an interest: we run the engine in production, and this is where it does and does not hold up. Upstream facts, including the issues cited, were checked against the MiroFish repository on 21 October 2026 (its documentation is unchanged since commit 39d8491).

The short answer

Good for: exploring how a public might react to a document, an announcement or a decision; finding objections, misreadings and factions early; and deciding what to ask real people.
Not proven for: predicting outcomes. The README reports no accuracy figure, and a proposal to score simulations against real outcomes (issue #813, September 2026) is still open.
Watch for: different conclusions from run to run and model to model, invented quotations in reports, and model costs that rise quickly with agents and rounds.

What MiroFish does well

It simulates a discussion, not a survey. The agents read the material and then react to each other over rounds, on simulated platforms modelled on Twitter and Reddit, so objections can spread and groups can split, which a one-shot answer cannot show.
It starts from your material. It builds a knowledge graph of the people and organisations your document mentions and turns them into agents, so the simulated public is about your situation rather than a generic panel.
You can question it afterwards. The last step of the workflow lets you chat with any agent and with the report agent.
It is open source. The code is published under the AGPL-3.0, so you can read how it works, run it yourself and change it.
It stands on published research. The simulation runs on OASIS, an open-source framework from the CAMEL-AI team, whose paper reports reproducing phenomena such as information spreading and group polarisation.

Where it falls short

No accuracy figures. The README reports none; its own financial and political prediction examples are listed as coming soon; and the backtesting proposal (#813) had no comments when we checked.
Results vary. An open issue (#17, January 2026) reports that the same material can lead to a different conclusion with a different model, and suggests running several simulations and keeping what they agree on.
The report can invent evidence. An open bug report (#529, April 2026) describes the report agent writing its own tool results, which can put made-up usernames, quotations and statistics into a report as if they came from the simulation.
It is one model playing many people. The agents are generated by the same language model, so its assumptions and blind spots repeat across the simulated public rather than cancelling out.
It costs more than it looks. The software is free, but every agent action is a model call. The README warns that consumption is high and suggests starting with fewer than 40 rounds, and Zep Cloud is free only up to its monthly quota.

How to get good results from it

Give it the real material, not a summary, and describe the people you want to understand in concrete terms.
Ask it to explore, not to predict: which objections come up, who disagrees with whom, and what changes over the rounds.
Run it more than once, ideally with more than one model, and rely on what the runs agree on.
Before you repeat a quotation from a report, find it in the agents' own posts.
Take the open questions to real people.

What a hosted service changes

MiroFish.co.uk is an independent hosted service built on the open-source MiroFish project. It runs a modified version of the engine for one job, stress-testing a decision, and packages it as a study. Two of the problems above are handled differently there: every quotation in a report is checked against what the simulated audience actually said, and shown with the round it came from, before the report is published; and the size of a run, and its price, are fixed by the study type, so there is no token bill to watch.

The other limits apply to any synthetic audience, ours included: the results are directional, not a forecast. See how to use MiroFish, or run one free Quick Study on a real decision.

Questions people ask

Is MiroFish accurate?

Nobody has shown how accurate it is. The project publishes no accuracy figures and no comparison of its simulations with real outcomes, so treat what it produces as scenarios to explore, not as predictions.

Can MiroFish predict the stock market?

There is no evidence that it can. The README lists financial prediction examples as coming soon, and nothing published compares its simulations with real market outcomes.

Why does MiroFish give different results each time?

Each run samples a language model, and the agents' choices build on each other over the rounds, so runs differ, and a different model can change the conclusion. Run it several times and rely on what the runs agree on.

Can I try MiroFish without installing it?

The upstream project has an online demo of a simulation it prepared. To run your own material without installing anything, use a hosted service; MiroFish.co.uk is one.

Sources

Start a free study