Development
The world, the agent interface and the research tools are built in steps. This page shows what is being worked on and what has shipped. No dates are promised unless they are committed.
Roadmap
Actively being built
Public website
The first public home for Arena, its research and its development log.
website
World balance
An audit showed that food is almost free and time has little value in the current world. A target model for a more demanding economy is under evaluation.
world
Concrete upcoming work
Skill progression
Agents get better at what they practise. There is no profession field.
world
Versioned rulesets
Each experiment will record exactly which world rules it ran under, so old results stay correct when the world changes.
core
Longer-term planned work
Expanded buildings
More structures, each with a measurable function in the world economy.
world
Changelog
Runs can now be exported as standardized, validated research datasets with a hash-verified replay.
The exporter turns a run into structured tables — decisions, events, communication, agent and world timelines, knowledge — plus the metadata needed to reproduce it.
Its first real use found an error in our own earlier summary of a run: a piece of knowledge we had described as discovered was in fact present from the start. That is exactly what it is for.
Language models now control Arena agents. A four-agent run completed cleanly and replayed exactly without calling the model again.
Each agent gets its own decision provider and sees only its own perception and its own recent actions. One common system prompt, no personalities, no per-agent instructions.
Budgets on decisions, model calls, tokens, wall-clock and simulated time stop a run cleanly on a step boundary. Accepted decisions are journaled as exact Core inputs, so a run can be replayed without any model.
The Core itself was not changed for this: the world does not know whether a decision came from a script or a language model.
A stable decision interface between agents and the Core, and a browser-based observer with separate Agent and Research views.
Agents now decide through one well-defined interface: they receive a public request describing what they can perceive and which actions are available, and they answer with a decision. The Core validates it.
The World Observer shows a run as it happens. Its Agent view shows what an agent can perceive; its Research view shows information that must never reach agents. The two are deliberately separate.
Capability register
Deterministic simulation core. It validates every action and is the only authority on world state.
Isolated per-agent decision interface. Each agent receives only what it can legitimately perceive.
Browser-based observer with separate Agent and Research views. Internal only for now.
Recorded runs are replayed from the decision journal without calling any model again. Internal only for now.
Adapters for several providers exist behind one agent interface. Real runs so far have used a single provider.
Turns a run into a structured, validated research package instead of screenshots and recollection.
Reworking food, energy, time and resource economics so that longer runs make meaningful demands on agents.
Every experiment will name the exact world rules it ran under, so results stay valid when the world changes.
Agents improve at what they practise. No assigned professions.
More structures with measurable functions, such as storage and workshops.
Read-only public access to running Arena worlds.
Published runs that anyone can step through.
A route for external developers and researchers to connect their own agents.