You paste a headline. It builds the people in it, lets them argue for ten minutes, and hands you a market question.
MiroShark writes the market question for you, and it gets there by a genuinely interest route.
You give it a scenario. A headline, a press release, a policy draft, whatever you've got. It reads the text, works out who actually has a stake in it, and casts them as agents. The people, the companies, the institutions named in your input, each with their own incentives and posting style, with public figures grounded against live web research so they reflect the current version of themselves.
Then it runs them. The agents post on a simulated Twitter, argue on a simulated Reddit, and trade an internal market, all at once and wired together. Traders see what's being posted, posters see where the price is. Someone says something inflammatory, the price moves, other agents notice and adjust. You get an actual narrative rather than a summary.
What comes out the other end, ten minutes later:
A sharp question. The hard part of creating a market has never been having the idea, it's the wording. Named actor, specific action, hard deadline, resolvable without argument. Get any of it wrong and you spend resolution day fighting in the comments about what counts. Real output, unedited, from public runs:
- "Will the Texas Attorney General's office file a formal lawsuit or injunction against Samsung..."
- "Will the university administration issue a formal statement modifying or delaying the income-blind policy..."
- "Will the AZZLE protocol's total active AI agent count exceed 500 by December 31, 2026..."
Named actor, specific action, hard deadline. That's the chore, done.
A catalyst map. Every run ranks which posts actually moved the room and which agent wrote them. One geopolitical run I read surfaced an energy-chokepoint threat, a military rehearsal, and a new alliance entering as a wildcard, ranked by how much they shifted the room. That's a watchlist for what to monitor, generated from a paragraph.
The disagreement, not just a consensus. It reports the split rather than collapsing to one answer. One run ended 44.4% bullish, 33.3% neutral, 22.2% bearish, with belief drift charted round by round so you can see the exact point the room fractured. A single price throws all of that away.
A report where everything has a receipt. Every claim links to the specific post or trade it came from. You can click through to the moment an agent changed its mind and read the argument that did it.
And you can push on it while it runs. DM any agent and it answers in character from its current belief state and posting history. Inject breaking news into a live timeline and watch the room absorb it. Or fork the run on a counterfactual and put both branches side by side, which is scenario work you'd otherwise be doing in your head.
It's open source, AGPL, around 1.4k stars, and self-hostable end to end including against local models.
Easiest way to judge it: 240 completed runs are public and free to read. No account, no payment.
Public runs: miroshark.xyz/sim
Repo: github.com/MiroShark/MiroShark