Skip to the article
DOTAFS

Personal learning notes

Paper note / 2026-09-15arXiv:2608.26081v122 min read

SwarmWorld: Artifacts as the Long-Term Memory of an Agent Society

SwarmWorld is an experimental platform for studying how LLM agents build technological culture through persistent artifacts, observation, modification, and reuse in a shared world. It asks a harder question than whether agents can converse: can participants with no assigned professions leave a technical inheritance that later participants genuinely use?

Paper
SwarmWorld: Stigmergic technological evolution in societies of language-model agents
Authors
Subhadeep Pal · Fiona Y. Wang · Markus J. Buehler
Institution / date
MIT · 2026-08-26
Code
Apache-2.0 simulator, traces, Observatory, replay tools, and human client

In one sentence: shared societies preserve a more disturbance-resistant portfolio of solutions, while isolated search can still discover the strongest individual artifact. The paper supports artifact-mediated social learning, not a universal claim that groups always beat individuals.

01 / RESEARCH QUESTION

From conversational memory to world state another agent can inherit

Many multi-agent systems place cooperation inside messages, plans, or short tasks. When the run ends, most of the useful work disappears with the context. SwarmWorld changes the medium of cooperation to persistent objects and executable controllers. Materials, structures, and programs remain in its two-dimensional world, where other agents can inspect, use, modify, and fork them.

This is stigmergy: participants influence one another through traces left in the environment rather than requiring continuous direct communication. The paper therefore measures whether a society develops a broader technological portfolio, whether artifacts evolve across authors, and whether the world remains useful under disturbances the agents did not see while building it.

Participants

50, 100, or 200 initially homogeneous agents, with no assigned roles or crafting recipes.

Shared medium

A spatial environment, resources, persistent artifacts, and executable controllers.

Evolution

Exploration, construction, testing, cross-author reuse, and forks produce an emergent division of labor.

Independent test

Agents are removed before a deterministic simulator applies eight held-out disturbance schedules.

What it adds to earlier approaches

ApproachCarrier of inheritanceCan successors modify it?SwarmWorld's distinction
Generative townsMemory, relationships, dialogueMostly changes social narrativeTreats executable artifacts as public memory
Best-of-N searchMutually isolated candidatesNo cross-agent inheritancePreserves authorship, use, and fork lineages
Role-assigned teamsTask plans and messagesWithin a predefined organizationRoles emerge from building, use, and maintenance
SwarmWorldPersistent artifacts in a shared worldUse, rewrite, combine, and forkTests the surviving portfolio under unseen disturbances
02 / METHOD MAP

What happens during one cycle of technological evolution

01 / SENSERead a local world sliceResources, nearby objects, memory, and work left by other agents
02 / PLANProduce a bounded planOne model, prompt, and action schema; at most 12 planned actions
03 / BUILDLeave an artifactResearch materials, construct objects, or write an executable controller
04 / REUSEFork across authorsUse, inspect, and modify other agents' work to create a lineage
05 / TESTRemove the buildersFreeze the world and evaluate it under eight unseen disturbances

The experiment keeps model capability fixed: gpt-5.6-luna, temperature 0.7, low reasoning, and the same prompt and action schema. The main grid runs for 800 ticks at N=50, 100, and 200, with four matched seeds per cell. Long-horizon runs use N=100 for 3,200 ticks.

“Culture” here is not a personality preset. It is an explicit mechanism for recording and transmitting experience. A shared-society condition without explicit culture helps separate the effect of the common environment from the effect of formal knowledge-sharing tools.

Evidence boundary: each experimental cell has only four paired seeds. The results support effect sizes and mechanism-level interpretation, not a stable population-wide “win rate” for swarms.

03 / EVIDENCE

The shared society wins on portfolios, not on every objective

LONG HORIZON3,200 ticks

N=100 runs expose cumulative development rather than a short burst of search.

NON-CREATOR USE>95%

More than 95% of artifacts were eventually used by an agent other than their creator.

LINEAGE9.75

Mean executable-lineage depth with explicit culture; the deepest path reached 12.

INTERACTIVE EVIDENCE LEDGER

Switch metrics to compare three organizational conditions at the same endpoint. Bar lengths are normalized within each metric and should not be compared across tabs.

Endpoint portfolio resilience is 0.2474 with explicit culture, 0.2365 without explicit culture, and 0.1794 for isolated search. Shared societies leave a collection that covers more disruptions rather than one single champion solution.

Change the objective to the strongest artifact and isolated search reaches 0.3488, above explicit culture at 0.2380. The no-explicit-culture condition also produces 7.00 validated inventions on average, compared with 5.75 under explicit culture. The evidence does not support “more communication is always better” or “groups always beat individuals.”

Explicit culture instead makes cumulative development more visible: at tick 3,200 it produces 277.5 artifacts on average versus 238.5 without explicit culture. Executable lineages are deeper, and about half of eligible forks cross an author boundary.

04 / CLAIMS & LIMITS

Separating paper evidence from project inference

PAPER CLAIM

Supported: shared worlds improve endpoint portfolio resilience; non-creators widely use artifacts, and executable work develops cross-author lineages.

OUR INFERENCE

Useful for an SAO-like world: objects, scripts, and use histories can form a more reliable social-memory layer than conversation summaries. Occupations can emerge from what residents build, use, and maintain.

NOT SHOWN

Still unproven: a playable 3D world, migration across releases, an economic loop, long-running service governance, or residents independently shipping entirely new game mechanics.

  • The physics is a surrogate. Normalized material properties support a game-level simulator; they are not an SI-calibrated physical predictor.
  • Knockout is not live recovery. Topological disturbances on a frozen world do not prove that a running society can detect, repair, and recover from failure.
  • The scale remains small. Fifty to 200 agents, at most 3,200 ticks, and four seeds are orders of magnitude away from an open world operating for months.
  • Strong individuals still matter. Isolated search leads on the strongest artifact. Diversity and peak performance are separate objectives.
05 / PROJECT MAPPING

For InfiniteAincrad: build an auditable inheritance chain first

WORLD STATE

Leave work inside the world

Give buildings, recipes, tool scripts, and quest templates stable IDs, authors, versions, dependencies, and use counts instead of leaving them only in resident context.

ACCEPTANCE

Let the environment test value

Before release, artifacts pass physics, performance, safety, and economic constraints. After release, real use, repair, and reuse determine whether they become infrastructure.

LINEAGE

Record who inherited whom

Every fork keeps its parent version, intent, and validation result. Emergent occupations then become observable through lineages and use—not declared by a prompt.

Why this is the primary reference

SwarmWorld already connects autonomous exploration, persistent work, successor use, and cross-author lineages into one measurable sequence. The first thing worth reproducing is not the 2D presentation; it is the artifact contract, lineage model, and post-agent acceptance test.