← BASELOAD English

Running Your First MiroFish Simulation — A Beginner's Guide [Part 3]

July 17, 2026 · 42 min read 한국어 원문

📚 Series — The Complete Guide to MiroFish ① Concept, history, use cases · ② System overview & install · ③ Beginner’s guide · ④ Scaling it up

The best way to get your first prediction out of MiroFish is to run one small, clear question through one small simulation. If you finished the install in Part 2, it’s time to open the web UI (localhost:3000) and actually run a virtual society. The bottom line first: success at the beginner stage rests not on fancy settings but on a good seed and a small scale. Feed a seed that carries a good question, set agents and rounds to a minimum to run end to end, then read the report as a story rather than a number — those three things are the whole of a first simulation. This guide takes a first-timer by the hand: prepare the seed → set the scale → run → read the report → god-view and agent chat → common mistakes.

Photo by Lukas Blazek on Unsplash

Step 1 — A Good ‘Seed’ Is 80% of the Result

MiroFish extracts a knowledge graph from the text you feed as a seed, then builds agents from it. In other words, the input is the raw material of the world. A thin seed makes a thin society; a rich, focused seed brings the simulation to life. The seed is the very first thing a beginner should care about.

A good seed has three qualities. First, it carries context — not just “Samsung stock outlook,” but the related news articles, reports, and background so that the actors and the stakes emerge. Second, it has a single focus — mixing many topics scatters the agents; the question should converge, like “how will opinion split after this policy announcement?” Third, it features people and groups — MiroFish’s power lies in “who reacts and how,” so the seed needs stakeholders with differing positions to make the graph rich.

A good first beginner seed looks like this: the full text of one or two controversial news articles, plus a one-line question — “predict how public opinion on this issue will split into factions and shift over the next two weeks.” Short, but people, stakes, and a question are all in it.

Step 2 — Agent Count and Rounds: Start Small, No Exceptions

Once the seed is in, set the simulation’s scale. The one rule a beginner must keep here: start small, always. As we saw in Part 2, MiroFish’s LLM calls explode in proportion to agent count × rounds, and that’s your cost. The repository itself says to “start with small simulations of fewer than 40 rounds.”

Source: MiroFish repository guidance (GitHub).

The goal of the first run isn’t an “accurate prediction” but “the pipeline running from start to finish.” Watching the whole process complete once — a knowledge graph built, agents generated, conversations exchanged on the two platforms, a report produced — that’s the success criterion for a first simulation. Scaling up comes after.

Step 3 — Run It and ‘Watch’

Once you start the simulation, the web UI shows agents posting and commenting across the two social platforms (Twitter-style and Reddit-style) and influencing each other, in real time. Here’s a pleasure and a key point beginners often miss: don’t just wait for the result — watch the process.

This visibility of process is exactly where MiroFish parts ways with other prediction tools. A statistical model spits out a single probability a few seconds later and it’s done. MiroFish shows you how opinion forms, right before your eyes. Which agent lights the first spark, which argument the crowd rallies around, the moment the tide turns — that trajectory itself is the first insight. As rounds progress, each agent’s memory updates and positions sometimes shift.

Step 4 — Read the Report as a ‘Story,’ Not a ‘Number’

When the simulation ends, a dedicated ReportAgent organizes the whole flow into a structured forecast report. Here’s the misreading beginners commit most often: pulling only the “one-line conclusion” from the report. MiroFish’s report wasn’t built to be read that way.

The right way to read it: First, see the terrain of factions — grasp how many groups opinion split into and each group’s core claim. Second, find the moments of change — which round the flow moved sharply, and what triggered that turn. Third, understand the conclusion as branches — MiroFish often shows “A under these conditions, B under those” rather than declaring “the answer is A.” The report is a map, not an answer key — it shows which roads you could take.

Step 5 — Dig In With God-View and Agent Chat

After reading the report, the real fun of MiroFish begins. There are two interaction tools.

One is the god-view. You can inject variables mid-simulation — say, insert a new event like “what if the government issued a rebuttal here?” and re-run to see how opinion changes. This is the “counterfactual scenario” experiment. Being able to branch multiple futures from one seed is MiroFish’s strength.

The other is the direct agent chat. Ask a specific virtual persona “why do you think that?” to check the individual logic beneath the collective result. It’s a way to interrogate the “why” that a poll’s results page never gives you. For a beginner, just posing a question to one or two representative agents from a faction that caught your eye makes the result far more three-dimensional.

Five Common Beginner Mistakes

Knowing the frequent first-simulation mistakes in advance saves time and money.

1. Seed too thin — feeding a single keyword and expecting a rich result. Feed text with context, people, and stakes.
2. Going big from the start — starting at 1,000 agents × 50 rounds and burning cost. Always complete a small run first.
3. Reading only the one-line conclusion — ignoring the report’s “process and branches” and hunting for an answer. MiroFish’s value is in the narrative.
4. Trusting a single run as the answer — it’s sensitive to initial conditions, so run the same seed several times and look at the tendency.
5. Ignoring API cost — running big without checking your LLM key’s usage and limits. Always monitor consumption on the dashboard.

Avoid just these five and the first experience is far smoother. In sum — small question, small scale, watch the process, read as branches, run multiple times.

So What — The Goal of a First Simulation Isn’t ‘the Answer’

The first thing a beginner should get from MiroFish isn’t a “correct prediction.” It’s “the sense of expanding your thinking by running a virtual society.” Branching multiple futures from one seed, watching factions form and the tide turn, asking individual agents “why” — learning to treat prediction not as a “one-line probability” but as an “explorable map.” Understanding, hands-on at the beginner stage, MiroFish’s essence from Part 1 — “a scenario lab, not an answer machine” — is the goal of Part 3.

In the next [Part 4], we’ll cover how to scale and apply what you’ve learned — API automation, running large simulations, real-world scenarios in markets, policy, and elections, and their limits. You’ve run your first simulation; now it’s time to forge this tool into a real weapon.

Frequently Asked Questions (MiroFish for Beginners)

Q1. What do I do first for a MiroFish simulation? Open the web UI (localhost:3000), feed text with context, people, and stakes as a “seed,” and attach one clear question. Set agent count and rounds to a minimum (under 40 rounds) and first confirm the pipeline runs end to end.

Q2. How many agents and rounds should I use? For a beginner’s first run, a few dozen agents and 10–20 rounds (under 40) are recommended. MiroFish’s LLM cost spikes with scale, so aim for “completion” over accuracy — start small, then grow.

Q3. How do I read the report? Don’t read only the one-line conclusion; look at ① how many factions opinion split into, ② where the flow changed, and ③ how the conclusion branches by condition. The MiroFish report is a “map” of possible roads, not an “answer key.”

Q4. What mistakes do beginners commonly make? Feeding too thin a seed, running big from the start and burning cost, reading only the conclusion, trusting a single run as the answer, and ignoring API usage. The principles: small question, small scale, watch the process, and repeat.

This article is a general usage explainer based on information published as of July 2026, and is not advice recommending investment decisions on any specific security or product.