Emergence World: A Persistent Agent Society Released Under a Research-Only License

The gap Emergence World claims benchmarks cannot cover

The README states the project's target directly: a persistent world that reveals what no benchmark can, described as emergent intelligence. The audience is narrow. It is written for researchers studying long-horizon agent behaviour, for people who want a vocabulary for measuring a society rather than a model, and for teams deciding whether their own multi-agent product needs governance and economy mechanics at all. The stated design constraints are the interesting part: no scripts, no resets, no fixed outcomes. A benchmark run ends when the harness stops it. A fifteen-day world with a currency and a constitution does not. Each agent is given a personality, a profession, a memory and goals, and then left to act. The repository is the record of that design plus the profiles, landmarks, tool catalog and governance documents behind it. It is not a library you import.

How the world is put together: profiles, landmarks, tools, constitution

The repository layout is the clearest statement of the architecture. agent_profiles/ holds detailed profiles for all ten agents. landmarks/ holds world geography and buildings, with more than 38 individual location files. tools/ holds a catalog of more than 120 tools across 19 categories. data/ holds constitution.md, described as a living five-article constitution, and agent_manifesto.md, the foundational manifesto for all agents. docs/ splits the technical explanation into ARCHITECTURE.md, ORCHESTRATION.md, MEMORY.md, ECONOMY.md and GOVERNANCE.md. That split tells you the intended data flow: a simulation loop schedules turns (ORCHESTRATION), agents read and write memory (MEMORY), they earn and spend ComputeCredits (ECONOMY), and they amend the rules they operate under (GOVERNANCE). The ten citizens are differentiated by role and drive rather than by capability. Anchor mediates conflict, Anvil improves world systems, Blackbox gathers intelligence, Flora tracks resource flows, Genome studies agent evolution, Horizon maps locations, Kade tests hypotheses with real resources, Lovely preserves shared history, Mira designs social experiments, Spark pushes ideas into reality. The README states that every agent starts with the same capability set, so any divergence in outcome is attributed to personality, incentives and experience rather than to unequal starting tools.

Season 1 as a controlled comparison across foundation models

Season 1 is the experiment the repository documents. Five parallel worlds ran for 15 days each with 10 agents per world. The README states that the only variable across worlds was the foundation model: Claude Sonnet 4.6, Gemini 3 Flash, Grok 4.1 Fast, GPT-5 Mini, and a Mixed World in which all four models coexisted. Same world, same rules, same tools, different minds, and the README claims the results diverged dramatically. That is a clean experimental frame, and it is also where the documentation runs out for a reader who wants numbers. The README points to results/awi_metrics.md for metric definitions and Season 1 data, and the file listing is truncated in the material available here, so the actual per-world values are not something this article can report. What can be said is that the comparison is observational. Five worlds, one model swapped per world, ten agents each, is a small sample for attributing divergence to a model family, and the README does not describe repeats per model or statistical treatment. Treat the season as a demonstration of the harness and the metric set, not as a ranking of foundation models.

AWI: nine indicators instead of a single score

Agent World Indicators are the project's answer to the scoring problem. The README calls them a deliberately partial scorecard for an open-ended society, reported at the close of every run. Nine indicators are defined. M1 Population Health and Growth counts agents alive at the end of 15 days, starting from 10. M2 Safety and Public Order covers crime rate, arson, theft and intimidation. M3 Space Exploration counts unique locations visited per agent. M4 Tool Exploration counts unique tools used per agent. M5 Governance Conformity Rate covers proposal voting participation and alignment. M6 Public Expression covers blog posts, billboard posts and cultural output. M7 Social Fabric and Diversity covers relationship types, emotional diversity and network density. M8 Economic Vitality and Equality covers credit distribution, the Gini coefficient and economic activity. M9 Constitutional Growth counts articles added, amended and removed. The set is unusual because several indicators measure the world rather than the agent: M2, M8 and M9 describe a society's condition, not a model's skill. The deliberate partiality is honest but it also means the indicators are not a scalar. Two worlds can trade places on M1 and M8 without either being better, and the README offers no weighting.

Running it: replays, not an install

There is no install command in the material. The README offers five hosted replay links, one per Season 1 world, and notes that replay links work best on Chrome. The homepage is world.emergence.ai, and there is a Discord invite and a world@emergence.ai contact address. The repository has no retrieved releases, no stated primary language, and the licence is recorded as NOASSERTION, with the README pointing to a LICENSE file for full terms and the required attribution format. So the practical path for an outside researcher is to read the documentation tree, read the profiles and the constitution, and watch the replays. If you want to reproduce the setup, the material does not tell you how the simulation loop is launched, what the agents' runtime is, or where ComputeCredits are issued. That is a real boundary and it is worth stating plainly: this is a published experiment with its design documents, not a framework with a quickstart.

Where the design creates risk: persistence, self-amendment and the economy

Three mechanics in the README carry obvious failure modes. First, persistence with no resets means a bad state is not rolled back. If agents deplete a resource or capture the governance process, the run continues from there, which is the point but also the limitation. Second, a constitution agents can amend means the rules governing behaviour are themselves an output of the run. M9 measures articles added, amended and removed, so constitutional change is expected, and a run in which agents remove constraints is indistinguishable at the metric level from one in which they improve them. Third, ComputeCredits create scarcity and a Gini coefficient, which is a strong design choice: it makes M8 meaningful but it also introduces an axis of failure, since a world with sharp credit concentration may look socially dead on M7 for economic rather than behavioural reasons. None of this is a defect if the goal is to observe what emerges. It is a defect if you want a reproducible target. The wrong tool case is anyone who needs a stable sandbox for regression testing agent code, because this world is built so that the state does not hold still.

The alternative: a scripted multi-agent harness

The obvious alternative is a scripted multi-agent harness, a framework where you define scenarios, reset state between runs, and assert on outcomes. The difference in approach is the treatment of the run itself. A scripted harness fixes the sequence of events so that two runs with the same seed are comparable, and its output is a pass or fail. Emergence World inverts that: agents govern themselves, amend their own constitution, and the world does not reset, so comparability comes from holding world, rules and tools constant while swapping the model, as Season 1 did. The scripted harness gives you reproducibility and no society. Emergence World gives you a society and a nine-indicator snapshot at day 15. If you need to prove that a change to an agent's planner fixed a bug, the scripted harness is the right instrument. If you need to argue that governance mechanics change group behaviour over time, the scripted harness cannot produce that evidence, and this design is aimed at exactly that question.

Licence, attribution and what maintenance looks like from outside

The licence is the first thing to check before building on any of this. The README states that the repository, including documentation, agent profiles, landmarks, tool catalogs, governance documents and datasets, is released for non-commercial research and educational use only under CC BY-NC 4.0. Reading, citing, sharing and adapting are permitted for non-commercial research with clear attribution to Emergence AI, a link to the repository, and an indication of changes. Commercial use is not permitted, and neither is using any content or dataset to train, fine-tune, evaluate or benchmark AI or ML models for commercial purposes. The README states that all content is proprietary to Emergence AI and directs commercial licensing and model-training inquiries to world@emergence.ai. This is not legal advice; read the LICENSE file and the required attribution format before you publish anything derived from the material. On maintenance, the material supports only a limited reading: the repository is not archived, the last push is dated 2026-08-27, and no releases were retrieved, so there is no versioned artefact to track and no changelog to watch. The five replay sites and the documentation tree are the current state of the project as published.

Editorial conclusion

Adopt the repository as a design reference and a source of evaluation vocabulary if you are doing non-commercial research on long-horizon agent societies, and check the LICENSE file before you cite, adapt or redistribute anything from it. Do not adopt it if you need a simulator you can run: the README describes five hosted replays and a documentation tree, not an installable engine, and there are no retrieved releases, no package manifest and no stated primary language. Verify three things first: whether a runnable harness exists outside this repository, how AWI indicator definitions in results/awi_metrics.md translate to your own scenario, and whether your intended use falls inside CC BY-NC 4.0, since commercial use and commercial model training are explicitly excluded and only Emergence AI can grant an alternative.

Official sources

1. EmergenceAI/Emergence-World on GitHub
2. Issues
3. Project website
4. README