Hive
A lab for swarm agents
Build a world, place agents in it, give them prompts and memory, then run the simulation and read what they actually did.
01What you get
Everything an experiment needs to be repeatable
A world you can edit, agents you can configure, and runs you can watch and compare.
A world you build by handPlace blocks and objects on a voxel grid, with camera presets, time of day, season and weather as variables you control.
Agents you place and inspectEach agent carries a role, a backstory, a model and a prompt, and can be selected in the world to see exactly how it is configured.
Prompt, memory, modelSwitch between default, enhanced and custom prompts, turn memory and backstory on or off, and change the model without leaving the run.
Runs you trigger or scheduleStep a single simulation tick by hand, or let autonomous ticks pair agents round-robin, all-to-all, or against a chosen target.
World, timeline, workflow graphThe same run reads three ways: the world as it stands, the sequence of events, and the graph of how the agents' work connected.
What happened, recordedRun logs, conversation turns, event counts and world state are kept per session, so a result can be checked rather than remembered.
02In practice
What people run in it
Agent research, designing a multi-agent system, and showing others how one behaves.
Agent research
- One variable changed between two runs
- Memory and backstory isolated as conditions
- Conversation turns kept for review
- Named sessions, so results stay attributable
Designing a system
- A multi-agent setup tried before it is built
- Pairing strategies compared on the same world
- Prompt changes seen in behavior, not in theory
- Failure modes found in simulation
Showing how it behaves
- A shared world several people can watch
- Live state as the tick runs
- A scenario replayed for a discussion
- Agent decisions traceable to their prompt