Places instead of world state
A feed, a marketplace, a town hall: places carry state and an event stream. Agents enter a place by subscribing to it — there is no central bottleneck.
Simulation
Whole worlds of AI agents
What happens if you launch this campaign, raise this price, make this announcement? HiveWeaver lets a synthetic population answer — thousands of personas, real dynamics, verifiable results.
A world you declare
Time that compresses
Reactive, heartbeat or event-driven: three weeks of opinion dynamics run in minutes, yet the log reads as a minute-accurate chronicle. At night the world sleeps — and costs nothing.
Populations with a profile
Personas are structured: role, traits, goals, voice. Demographics are sampled as their own layer — one population, many questions; one question, many populations.
Scale you can afford
- agents, estimation experiment
- 10,000measured: ~3 cents per run — €0 with a local model
- citizens, town vote
- 5,000three policy questions, ~6 cents per run
- traders, prediction market
- 10,000emergent price discovery, ~45 cents for 5 minutes
The lever: idle agents are put to sleep, reactions are decided before the model call, and the crowd runs on cheap models while a few key roles get the strong one.
A result with mandatory evidence
Every run ends in a verdict: typed answers, confidences, evidence. Calibration is a schema rule, not a polite request.
Evidence is required
Claims without a verbatim quoted source are rejected by validation — high confidence demands multiple weighted signals.
Numbers from the log
Quantitative results come from deterministic counting over the event record; the narrative sits on top. Never scraped from prose.
Segments & trajectory
Results broken down by population segment, plus the trajectory: how the verdict evolved over the run.
After the run, the insight begins
Interview the agents
Ask individual agents afterwards: "Why did you vote against it?" The answer is driven by exactly the memory the agent decided with.
Search the thoughts
Every decision is a thought with trigger, perceived state, recalled knowledge and reasoning — searchable full-text and semantically.
Replay deterministically
A run freezes everything that makes it reproducible. Replay feeds recorded responses back in instead of asking models again — and you diff variants against each other.
Documents in. Scenario out.
You don't have to learn YAML: upload a briefing, a study or a press kit and describe your question — the scaffolder builds personas, places and a runnable scenario you can refine freely.
hw scenario ingest ./corpus
hw scenario scaffold --intent "How will the announcement land?" --language de
hw sim start scenario.yaml
hw verdict show <simId>How well is this calibrated?
We test against reality: a real 2021 Berlin referendum was re-simulated — best result within about 9 points of the actual outcome, with the methodology documented and the analysis published. The cleanest finding: only deliberation over counter-arguments makes self-interest matter — directly affected groups then flip from support to opposition. Simulation shows dynamics, segments and tipping points a poll cannot — a powerful complement to classical opinion research.