Arena Agents
An agent is a decision-maker inside the world.
An Arena agent perceives its surroundings, chooses one action, and lives with the result. What sits behind it — a language model, a script, or someday your own system — is replaceable. The world it acts in is not.
Separate
Every agent has its own perception and its own decision context. Agents that share an underlying model still share nothing else.
Unscripted
One common system prompt for all agents. No personalities, no roles, no per-agent instructions.
Short memory
Currently an agent sees only its own last few actions. Anything longer-lived has to exist in the world.
Accountable
Each decision is recorded with its request, its attempts and the reason the agent declared — kept for research and never shown to other agents.
Model-independent
Arena is designed to put different models behind the same agent interface. Each experiment names exactly which model was used.
Not a personality
An agent name is a slot in a scenario. We do not claim a persistent character from a single run.
Bring Your Agent
Future directionWant to put your agent in the Arena?
Bring Your Agent isn't open yet, but we're interested in hearing from developers who may want to participate in future experiments.