Pith. sign in

REVIEW 2 major objections 5 minor 18 references

Amorphous Fortress Online: Collaboratively Designing Open-Ended Multi-Agent AI and Game Environments

T0 review · 2 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Amorphous Fortress Online is a browser platform where users design, play, and share small multi-agent worlds built from simple finite-state machines, building toward a community database of open-ended AI environments.

desk verdict Genuine browser-based FSM environment platform with honest pilot data, but the 'diverse database' claim is ahead of the evidence. read the letter →

arxiv 2502.05632 v1 pith:BSPNITKS submitted 2025-02-08 cs.AI

M Charity , Mayu Wilson , Steven Lee , Dipika Rajesh , Sam Earle , Julian Togelius This is my paper · ORCID
classification cs.AI
keywords multi-agentsystemsfinite-statemachinesopen-endednessuser-generatedcontentcollaborativedesignartificiallifewebplatformemergentbehavior
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Amorphous Fortress Online is a browser-based platform where anyone can design, play, and share small petri-dish-like worlds made of simple finite-state machine agents. The paper argues that such a community platform can grow into a database of thematically diverse AI and game environments, with emergent behavior arising from transparent FSM logic that users can inspect while the simulation runs. If the platform gains a user base, it would give AI researchers a shared, remixable collection of small open-ended environments for training agents and generative models. A short study of the first release shows modest uptake: 33 fortresses total, only five from non-authors, alongside 174 plays.

What carries the argument

The load-bearing mechanism is the entity FSM: every entity class is an ASCII character plus a finite-state machine made of action nodes and condition edges, evaluated each timestep against the fortress grid. Thirteen node types cover autonomous behaviors (move, chase, take, add, clone, transform, and others) and player-controlled actions, while edges trigger on no condition, a step count, proximity within a distance, adjacency, or contact. This FSM formulation is simple enough for users to design in a drag-and-drop canvas yet expressive enough to reproduce game interactions such as a Bokoblin chasing Link when within five spaces. The engine's termination conditions (extinction, overpopulation, inactivity) bound every simulation and let users observe emergent behavior in a compact setting.

What would settle it

Track all fortress submissions for one year after the public release: if the share of non-author fortresses stays below ten percent, or if submitted fortresses reuse essentially the same few entity themes and remix chains never exceed depth four, then the claim that a thematically diverse community database is forming would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that a carefully designed web interface can turn a simple FSM-based simulation engine into a collaborative repository of open-ended AI environments. Each fortress consists of a 16-by-8 grid of entities whose behaviors are defined by action nodes (idle, move, die, clone, push, take, chase, add, transform, move wall, plus three player-controlled variants) and condition edges (none, step, within, nextTo, touch). Users can write these definitions through drag-and-drop graph editing, a fortress placement screen, or a text editor with a validating compiler, then play the simulation and see the active FSM states highlighted in real time via the X-ray feature. The authors present this working platform as the foundation for future work in training reinforcement learning agents and generative models on the submitted fortress data.

Load-bearing premise

The platform's value as a diverse environment database depends on outside users actually submitting many varied fortresses; the paper's own study found only five non-author fortresses among 33, so this adoption premise is not yet supported.

Editorial extensions

If this is right

  • Entities saved to a backpack can be dropped into a new fortress, with dangling character references randomly replaced, which supports cross-environment remixing and long design lineages.
  • The X-ray feature shows active nodes and recently used edges during play, making agent behavior inspectable and helping designers debug before submission.
  • The database of fortress definitions — layouts, entity FSMs, and submitted notes — is intended to serve as training data for reinforcement learning agents, recommendation systems, and generative models.
  • The node-frequency data reported (298 nodes across 143 entities) gives an early quantitative picture of which behaviors designers choose, a pattern the platform could track over time.
  • Remix depth, already reaching four in the short study, provides a measurable indicator of community reuse that future releases can grow.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next test is whether remix chains and backpack reuse predict the emergence of genuinely new behaviors; the platform's submission logs would allow that analysis without new infrastructure.
  • The gap between 174 plays and five non-author submissions suggests that a mixed-initiative generator that proposes a starting fortress may matter more than editor polish for growing the database.
  • Because each fortress is a compact, parseable artifact, the corpus could double as an open-endedness benchmark: one could measure how often submitted entities use new combinations of node types or produce long extinction-free runs.
  • The planned sprite-generation step implies that, if adoption grows, the site also becomes a continuous source of labeled small images paired with gameplay data, extending its value beyond the game AI community.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces Amorphous Fortress Online, a web-based platform for designing, playing, and sharing small multi-agent environments ("fortresses") driven by finite-state machine (FSM) agents. The system includes an entity editor, fortress editor, text editor with compiler validation, a play screen with an X-ray debugging feature, and a sharing/remixing pipeline with a "backpack" for saving entities. The authors describe the engine's 13 action nodes and five edge conditions, then report a short post-launch study: 174 plays, 33 fortresses, 143 entities, and 298 nodes. The abstract claims the system will provide a database of thematically diverse AI and game environments and lays groundwork for future AI training and generative modeling experiments.

Significance. If the platform were to attract a sustained user community, it could provide a valuable, heterogeneous corpus of simple, transparent multi-agent environments, and the X-ray explainability feature is a nice design choice for human-understandable agent behavior. The paper reports a working, open-source system and honest usage statistics rather than hiding the limited external adoption. However, the central claim of delivering a "thematically diverse" database is not supported by the presented data: the corpus is small, mostly author-made, and no diversity measure is given. The significance is therefore currently prospective rather than demonstrated; the contribution is a systems paper with a credible foundation, not yet a validated data-resource paper.

major comments (2)
  1. [Abstract and Section IV] The abstract states that the system "serves to provide a database of thematically diverse AI and game environments," but the data reported in Section IV cannot support this claim. The study found only 33 fortresses total, with the text indicating that six authors submitted 28 of them and only five external fortresses arrived after launch. No content analysis, genre/topic labels, or diversity metric is reported; the node-frequency histogram in Fig. 11 addresses node-type coverage, not thematic diversity. The authors should either reframe this claim as an intended future outcome (as the Conclusion already does) or provide an operationalized definition of thematic diversity and demonstrate that the existing corpus satisfies it.
  2. [Section IV] The usage statistics are internally inconsistent: the text first reports that six authors submitted 33 fortresses and that "only five additional fortresses were submitted after the release of the site," which implies 28 author fortresses, but the following sentence says "Outside of the author's submission of 21 fortresses, users submitted an average of two fortresses." This contradicts the earlier counts and also conflicts with the statement that there were five external submissions. Please correct the numbers and clarify the exact split between author and non-author fortresses, as well as the average external submissions.
minor comments (5)
  1. [Section III.A and Section III.D] The engine description in Section III.A says the fortress grid is 16 x 8 spaces, while Section III.D states that the fortress editor uses a "14 by 6 grid space area." Since the text-editor compiler checks fortress dimensions, this discrepancy is confusing and should be aligned.
  2. [Section IV] The sentence "There were a total of 143 entities made in the website and a total of 298 nodes" is ambiguous: does it count distinct entity definitions or placed instances, and does it include author submissions? Please clarify the counting unit.
  3. [Section III.F] In the X-ray paragraph, "transtitioned" should be "transitioned" (Figure 8 caption area).
  4. [Figure 11] The histogram has no descriptive caption clarifying whether it counts nodes in all fortresses or per-entity; a short caption would improve interpretability.
  5. [Section IV] The phrase "The stark difference of fortress plays vs. fortress creation" would be clearer as something like "The contrast between the number of plays and the number of submissions."

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper is a systems report whose central database-diversity claim is empirically under-supported but not derived from its own inputs.

full rationale

Amorphous Fortress Online is a software and community-platform paper. Its central claim—'This system serves to provide a database of thematically diverse AI and game environments'—is a stated design goal, not a result derived from a fitted parameter, equation, or self-cited theorem. The engine is described as a JavaScript port of the authors' prior Amorphous Fortress engine [14, 15]; this is engineering lineage, not a load-bearing citation of an unverified result. Section IV's honest admission that only five non-author fortresses were submitted is a limitation on the database-diversity claim, but it is an empirical-evidence gap, not circularity: the claim does not reduce to its own inputs by construction. No predictions are made, no parameters are fitted and renamed as predictions, and no alternative hypotheses are ruled out by a self-cited uniqueness theorem. The paper is self-contained as a systems description and usage report, so no circular step is identified.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The platform introduces no new theoretical entities or physical postulates. The FSM action nodes, edge conditions, and fortress termination criteria are design choices from prior work or ad hoc engine constants. The central contribution is software infrastructure, so the axiom ledger is short and consists mainly of domain assumptions about the expressiveness of the FSM vocabulary and the viability of a user-generated dataset.

free parameters (2)
  • overpopulation threshold = 168
    Hand-selected engine constant that stops simulations with too many entities; not fitted to data and not central to the platform's claim.
  • inactivity threshold = 100 timesteps
    Hand-selected termination condition; arbitrary design choice in the engine.
assumptions (3)
  • domain assumption Finite-state machines with the 13 defined action nodes and 5 edge conditions are expressive enough to model a useful range of multi-agent interactions and games.
    This is the core modeling premise inherited from prior Amorphous Fortress work; Section III-A defines the vocabulary and the Zelda examples illustrate it, but no expressiveness proof or benchmark is given.
  • domain assumption Random direction choices in movement and chase actions are sufficient to produce interesting emergent behavior.
    In the engine, entities move in random directions unless chasing; the paper assumes this stochastic behavior leads to emergent interactions, asserted rather than demonstrated in Section III-A.
  • domain assumption Community-generated fortress data will be suitable for training generative and recommendation AI systems.
    Section V frames future AI training on user submissions as a goal, but the small study provides no evidence that the data distribution will support such training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Amorphous Fortress Online: Collaboratively Designing Open-Ended Multi-Agent AI and Game Environments." pith.science (2026). https://pith.science/paper/BSPNITKS

@misc{pith2026250205632,
  author       = {Pith},
  title        = {Pith review of: Amorphous Fortress Online: Collaboratively Designing Open-Ended Multi-Agent AI and Game Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BSPNITKS}},
  note         = {Machine review of arXiv:2502.05632}
}
read the original abstract

This work introduces Amorphous Fortress Online -- a web-based platform where users can design petri-dish-like environments and games consisting of multi-agent AI characters. Users can play, create, and share artificial life and game environments made up of microscopic but transparent finite-state machine agents that interact with each other. The website features multiple interactive editors and accessible settings to view the multi-agent interactions directly from the browser. This system serves to provide a database of thematically diverse AI and game environments that use the emergent behaviors of simple AI agents.

Figures

Figures reproduced from arXiv: 2502.05632 by the authors.

Figure 1
Figure 1. The community-based ZZT game design process show [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A visualization of an entity’s behavior FSM structure [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Two example agent interactions from Zelda: Tears of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: The fortress editor features a canvas window (B) with [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 4
Figure 4. Figure 4: The entity editor screen for Amorphous Fortress [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: The text editor screen for Amorphous Fortress Online [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: The fortress editor screen for Amorphous Fortress [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 8
Figure 8. Figure 8: The X-Ray feature (right) highlighting the Knight entity [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 7
Figure 7. Figure 7: The fortress play screen for Amorphous Fortress Online [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 9
Figure 9. Figure 9: Main screen of Amorphous Fortress Online with fortress selector and viewer [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: When logged in, a user can save an entity from one [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Frequency of entity action nodes a depth of 4 fortresses. Five fortresses with a play move, play push, or play move wall were made. V. REFLECTION AND FUTURE WORK Amorphous Fortress Online is intended to be a system to lay the foundation for future generative and recom…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

  1. [1]

    Open problems in artificial life,

    M. A. Bedau, J. S. McCaskill, N. H. Packard, S. Rasmussen, C. Adami, D. G. Green, T. Ikegami, K. Kaneko, and T. S. Ray, “Open problems in artificial life,” Artificial life, vol. 6, no. 4, pp. 363–376, 2000

    work page 2000

  2. [2]

    Open-endedness: The last grand challenge you’ve never heard of,

    K. O. Stanley, J. Lehman, and L. Soros, “Open-endedness: The last grand challenge you’ve never heard of,” While open-endedness could be a force for discovering intelligence, it could also be a component of AI itself, 2017

    work page 2017

  3. [3]

    Using Fractal Neural Networks to Play SimCity 1 and Conway's Game of Life at Variable Scales

    S. Earle, “Using fractal neural networks to play simcity 1 and conway’s game of life at variable scales,” arXiv preprint arXiv:2002.03896, 2020

    work page Pith review arXiv 2002

  4. [4]

    The neu- ral mmo platform for massively multiagent research,

    J. Suarez, Y . Du, C. Zhu, I. Mordatch, and P. Isola, “The neu- ral mmo platform for massively multiagent research,” arXiv preprint arXiv:2110.07594, 2021

    arXiv 2021

  5. [5]

    Picbreeder: evolving pictures collaboratively online,

    J. Secretan, N. Beato, D. B. D Ambrosio, A. Rodriguez, A. Campbell, and K. O. Stanley, “Picbreeder: evolving pictures collaboratively online,” in Proceedings of the SIGCHI conference on human factors in computing systems, pp. 1759–1768, 2008

    work page 2008

  6. [6]

    Baba is y’all: Collaborative mixed-initiative level design,

    M. Charity, A. Khalifa, and J. Togelius, “Baba is y’all: Collaborative mixed-initiative level design,” in 2020 IEEE Conference on Games (CoG), pp. 542–549, IEEE, 2020

    work page 2020

  7. [7]

    Aesthetic bot: interactively evolving game maps on twitter,

    M. Charity and J. Togelius, “Aesthetic bot: interactively evolving game maps on twitter,” in Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment , vol. 18, pp. 18–25, 2022

    work page 2022

  8. [8]

    Anthropy, ZZT: Boss Fight Books# 3, vol

    A. Anthropy, ZZT: Boss Fight Books# 3, vol. 3. Boss Fight Books, 2014

    work page 2014

Show all 18 references
  1. [9]

    Littlebigplanet

    MediaMolecule, “Littlebigplanet.” [PlayStation3], 2008

    2008

  2. [10]

    Littlebigplanet 2

    MediaMolecule, “Littlebigplanet 2.” [PlayStation3], 2011

    2011

  3. [11]

    Tiny online game engines,

    J. Warren, “Tiny online game engines,” in 2019 IEEE Games, Enter- tainment, Media Conference (GEM) , pp. 1–7, IEEE, 2019

    2019

  4. [12]

    The angelina videogame design system—part i,

    M. Cook, S. Colton, and J. Gow, “The angelina videogame design system—part i,” IEEE Transactions on Computational Intelligence and AI in Games , vol. 9, no. 2, pp. 192–203, 2016

    2016

  5. [13]

    The angelina videogame design system—part ii,

    M. Cook, S. Colton, and J. Gow, “The angelina videogame design system—part ii,” IEEE Transactions on Computational Intelligence and AI in Games , vol. 9, no. 3, pp. 254–266, 2016

    2016

  6. [14]

    Amorphous fortress: Observing emergent behavior in multi-agent fsms,

    M. Charity, D. Rajesh, S. Earle, and J. Togelius, “Amorphous fortress: Observing emergent behavior in multi-agent fsms,” ALIFE for and from games Workshop, 2023

    2023

  7. [15]

    Quality diversity in the amorphous fortress: Evolving for complexity in 0-player games,

    S. Earle, M. Charity, J. Togelius, and D. Rajesh, “Quality diversity in the amorphous fortress: Evolving for complexity in 0-player games,” in Second Agent Learning in Open-Endedness Workshop , 2023

    2023

  8. [16]

    G. N. Yannakakis and J. Togelius, Artificial Intelligence and Games . Springer, 2018

    2018

  9. [17]

    The legend of zelda: Tears of the kingdom

    Nintendo, “The legend of zelda: Tears of the kingdom.” [Nintendo Switch], 2023

    2023

  10. [18]

    The five-dollar model: generating game maps and sprites from sentence embeddings,

    T. Merino, R. Negri, D. Rajesh, M. Charity, and J. Togelius, “The five-dollar model: generating game maps and sprites from sentence embeddings,” in Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, vol. 19, pp. 107–115, 2023

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.