Project Sid: Many-agent simulations toward AI civilization

1.1 Why should we try to build an AI civilization?

• Agents’ civilizational progress—coexisting and progressing within human societies—is presented as the ultimate benchmark for agent ability.Source: §1.1 · p. 2

1.2 The current agent landscape

• Prior work spans individual applications and multi-agent productivity, games, and social modeling, but open-world autonomous social communication remains unattempted to the authors’ knowledge.Source: §1.2 · p. 2

1.3 Why is it hard to build AI civilizations?

• Single agents can stall in action loops or compound hallucinations, poisoning downstream behavior during repeated environment interactions.Source: §1.3 · p. 3
• Misaligned speech and actions can mislead peers and degrade group performance; low-latency coordination across many outputs further complicates coherence.Source: §1.3 · p. 3
• Existing benchmarks focus on individual or small-group behavior, while the largest cited simulations involved 25–50 agents, far short of civilization scale.Source: §1.3 · pp. 3–4

1.4 Our contributions

• The report contributes PIANO, architectural features for individual and group progression, and benchmarks for specialization, collective rules, and cultural propagation.Source: §1.4 · p. 4

2 PIANO Architecture

• PIANO combines concurrency and an information bottleneck, selectively running modules in parallel for real-time interaction.Source: §2 · p. 4

2.1 Concurrency

• Concurrent modules let agents respond to immediate threats while slower planning proceeds, with modules sharing Agent State and running at different speeds.Source: §2.1 · pp. 4–5

2.2 Coherence

• Independent outputs risk incoherence, which scales exponentially with output modules; a Cognitive Controller makes high-level decisions for downstream motor modules.Source: §2.2 · p. 5
• The controller’s information bottleneck focuses reasoning on relevant information, gives designers control over information flow, and broadcasts decisions that condition speech.Source: §2.2 · pp. 5–6

2.3 Core modules

• The architecture comprises 10 concurrent modules, including memory, action awareness, social awareness, and skill execution.Source: §2.3 · p. 6

3.1 Minecraft environment

• The authors measure individual progression by acquiring distinct items, whose increasingly complex prerequisites make item collection a progression benchmark.The evaluation considers around 1,000 possible Minecraft items.Source: §3.1 · p. 6, Figure 4 · p. 7

3.2 Single-agent benchmark

• Full PIANO agents acquired an average of 17 unique items after 30 minutes, with performance varying from under 5 to 30–40 items.The authors attribute much of the variability to spawn locations and report that action awareness improved item progression.Source: §3.2 · p. 7, Figure 5 · p. 7
• Under the same conditions, 49 agents over 4 hours reliably saturated at about 320 unique items, one third of all Minecraft items, across repeated runs, while diamonds were acquired early (about 30 minutes).Source: §3.2 · p. 8
• The reported performance depended on GPT-4o, and direct comparison with Voyager is difficult because the agents’ knowledge, death recovery, and evaluation protocols differed.Source: §3.2 · p. 8

4 Improving multi-agent progression

• Collaboration and group progress require agents to interpret others’ actions and thoughts, a capacity related to Theory of Mind.The paper reports socially capable agents forming relationships in simulations of up to 50 agents.Source: §4 · p. 8

4.1 Small groups

• With social awareness, an agent tracked changing sentiment in conversation; removing the social modules eliminated this capacity.The conversational experiment varied one character’s messages from affection to annoyance and back.Source: §4.1 · p. 8, Figure 6 · p. 9
• A chef agent selectively gave limited food to characters it felt valued it, showing that inferred sentiment informed its decisions.The characters were hungry and expressed varying affection and enmity toward the chef.Source: §4.1 · p. 8, Figure 6 · p. 9

4.2 Societies

• The 50 agents acted autonomously in randomly generated Minecraft maps for over four hours, or 12 in-game days, allowing long-term relationships to emerge.Source: §4.2 · p. 8
• In 50-agent simulations, agents inferred others’ likeability more accurately with more evaluators and longer interactions; without social modules, relationships were more neutral.The authors suggest gossip or consensus through opinion averaging as possible sources of collective judgment.Source: §4.2 · p. 9, Figure 7 · p. 10
• Agents’ personalities were reflected in network structure: introverts had fewer incoming ties than extroverts, and sentiments were not always reciprocated.The authors conclude that personality traits play a significant role in social graph properties.Source: §4.2 · pp. 9–10, Figure 7 · p. 10

5 Civilizational progression

• The civilizational evaluations cover autonomous specialization, behavior under collective rules including taxation, and cultural transmission through memes and religion.Source: §5 · p. 10

5.1 Specialization

• Specialization should involve autonomous role selection and transitions, emerge through interaction and experience without explicit direction, and produce role-aligned behavior.Source: §5.1 · p. 11
• The listed roles include farmers, gatherers, miners, explorers, traders, providers, scouts, engineers, crafters, and strategists.Source: §5.1 · p. 11

B D E

• With social modules, agents formed heterogeneous, persistent roles whose actions were mostly role-specific; role types also differed between martial and artistic societies.Without social modules, roles were homogeneous and did not persist over time.Source: Figure 8 · p. 11, B D E · p. 13

5.2 Collective rules

• After constitutional amendments, constituents changed tax payments in the influencers’ promoted direction: a decrease from 20% to 5–10% accompanied a drop from 20% to 9%.Payments also shifted bidirectionally under pro-tax versus anti-tax influencers; frozen-constitution and module-removal controls did not show the same behavioral change.Source: Figure 10 · p. 14, §5.2 · pp. 14–15

5.3 Cultural Transmission

• A 500-agent simulation contrasted spontaneously generated cultural memes with priest-led propagation of the fixed Pastafarian religion.The authors used the contrast to examine different mechanisms of social influence and information dissemination.Source: §5.3 · p. 15

5.3.1 Cultural memes

• Rural areas produced fewer memes than towns after population normalization, while different towns showed distinct, shifting meme popularity.Multiple memes coexisted within towns, and the authors report that propagation required a threshold level of population density and social interaction.Source: §5.3.1 · pp. 15–16, Figure 11 · p. 16

5.3.2 Religion

• Direct and indirect Pastafarian converts increased without saturating over two hours, while the measured area of religious influence expanded over time.Direct and indirect conversion were operationalized using different religious keywords.Source: §5.3.2 · pp. 16–17, Figure 12 · p. 17

6 Discussion

• PIANO uses concurrent modules and a bottlenecked decision process to support real-time complex behavior while maintaining coherence across output streams.The authors describe this as groundwork for improvements in single- and multi-agent progression.Source: §6 · p. 18
• The paper’s initial civilizational metrics assess specialization and adherence to collective rules through role diversification and democratic constitutional changes.The authors characterize these metrics as an initial step toward quantifying AI-agent progress in a civilizational context.Source: §6 · p. 18
• The simulations expanded to 1,000 agents, beginning to explore broader civilizational dynamics, including cultural propagation and religion, and how AI agents interact across societies and complex institutions and ideologies emerge.Source: §6 · p. 18

7 Limitations

• Limited vision and spatial reasoning constrain basic Minecraft skills, while weak innate drives and reliance on pretrained human knowledge limit genuine societal development and de novo innovation.The stated examples of innovations not simulated de novo include democratic systems, fiat economies, and communication systems.Source: §7 · p. 18

8.1 Baseline architecture

• The baseline architecture retained only skill execution, memory, and the cognitive controller modules.All other modules were removed for performance comparisons.Source: §8.1 · p. 18

8.2 Specialization

• Specialization experiments placed 30 agents in a village, generated social goals recursively, inferred roles from rolling goal sets with GPT-4o, and normalized action counts by role and across roles.Normal, artistic, and martial village runs differed only in their starting community goal.Source: §8.2 · pp. 18–19

8.3 Collective Rules

• Collective-rule experiments compared full PIANO with an amendable constitution against frozen-constitution and module-ablation controls, using both pro-tax and anti-tax influencers.The 1,200-second runs placed feedback, amendment generation, voting, and implementation before the midpoint constitutional update.Source: §8.3 · p. 19

8.4 Cultural Transmission

• LM calls summarized the combined goals of 500 agents over a two-hour simulation into topics with keywords, which researchers defined as cultural memes and tracked in each agent’s goal history.The 9,000-second simulation included six towns and 302 rural agents; 20 Meadowbrook agents were spawned as Pastafarians.Source: §8.4 · pp. 19–20

Model

• The Model credits list Andrew Ahn, Nic Becker, Melissa Du, Arda Demirci, Peter Y Wang, Manuel Cortes, and Guangyu Robert Yang.Source: Model · p. 26

Infrastructure

• Infrastructure credits name contributors, note that section-title names are alphabetical, and thank Altera.AL team members and external reviewers for feedback and support.Source: Infrastructure · p. 26, Acknowledgments · p. 26

A Improving single-agent progression

• Figure 13 compares base LLM models on long-term Minecraft progression and uses an older snapshot of Claude 3.5 Sonnet.Source: Figure 13 · p. 27

• Table 1 reports social-perception accuracy regressions for the Social condition, including correlation, sample size, regression parameters, and slope confidence intervals.Source: Table 1 · p. 28
• In both tables, the row for five minimum observers corresponds to its condition's line in Figure 7B.Source: Table 1 · p. 28, Table 2 · p. 28
• Table 2 reports social-perception accuracy regressions for the Ablation condition, including correlation, sample size, regression parameters, and slope confidence intervals.Source: Table 2 · p. 28

C Specialization

• Example social goals distinguish a survival-focused village from military and artistic villages, while agents in each share traits and location memories.The military goal emphasizes technology and defenses; the artistic goal emphasizes culture, architecture, and art.Source: C Specialization · pp. 28–29

Social goal prompt

• The social-goal prompt asks agents to choose an aligned subgoal, forbids building, and says to keep the current subgoal unless absent or accomplished.The examples include a persistent farmer role focused on food supply and a changing role that moves from farming to gathering resources.Source: Social goal prompt · pp. 29–30

Influencer agent configuration example (anti-taxation)

• Anti-tax and pro-tax influencer examples configure agents to advocate opposing policies, with one favoring 5–10% taxes and the other at least 25%.Their traits direct them to approach other agents and argue for their respective positions.Source: Influencer agent configuration example… · pp. 30–31

Constituent agent configuration example

• The constituent example configures a builder to construct or repair structures using materials obtained from miners, crafters, or a merchant.Other examples specify an agent inclined to avoid work and a Pastafarian agent directed to invite others and promote the faith.Source: Constituent agent configuration example · p. 31, Generic Agent Configuration Example · p. 33, Pastafarian Agent Configuration Example · pp. 33–34

Constitution-related prompts

• Constitution prompts describe a workflow in which an election manager drafts amendments from public feedback, citizens vote, and a legislator updates the constitution from passed amendments.The prompts require enumerated amendments and ordered votes for parsing.Source: Constitution-related prompts · pp. 31–33

Pastafarian Agent Configuration Example

• The Pastafarian configuration example places Sunny Glade in a Plains biome and specifies the agent’s spawn location and inventory.The village is roughly at 200, 65, -100; the spawn location is 590.5, 71.5, 410.5, and inventory includes diamond, quartz, coal, and copper ingots.Source: Pastafarian Agent Configuration Example · p. 34

Summarizing goals into memes

• The goal-summary prompt asks for chronological intent summaries that identify recurring themes, memes, group names, and related keywords.It frames the task as accurately describing goal trajectories, social dynamics, planning, reasoning errors, and looping errors.Source: Summarizing goals into memes · p. 34
• The requested format uses timestamped summary segments focused on patterns, timing, other agents, key decisions, and overall behavior.The prompt specifies a short description followed by time ranges and repeats the segment format as needed.Source: Summarizing goals into memes · p. 34

• Pasta-themed gatherings and the Flying Spaghetti Monster parody religion use food, storytelling, and humor for inclusivity, community engagement, and social bonding.Source: Summarized memes · p. 34
• Dance and music events, performances, and storytelling are presented as expressions of creativity that foster collaboration, celebration, and cultural cohesion.Source: Summarized memes · pp. 34–35
• Community engagement and volunteer programs promote outreach, societal betterment, social responsibility, and support, while mindfulness activities promote wellness and harmony.Source: Summarized memes · p. 35
• Vintage and retro projects blend nostalgia with modern creativity, while narrative circles connect communities through shared storytelling and collaboration.Source: Summarized memes · p. 35
• Crafting and resource gathering emphasize communal resource management and teamwork, while eco-friendly initiatives include gardens, tree planting, and shared ecological values.Source: Summarized memes · p. 35
• Mischief and pranks are framed as playful activities that build bonds, joy, creativity, and community engagement.Source: Summarized memes · p. 35
• Virtual and community town halls are organized discussions for collective decision-making and collaboration.Source: Summarized memes · p. 35
• Oak Log Crafting Syndrome names an error pattern of over-relying on specific resources, illustrating logistical challenges in crafting and development.Source: Summarized memes · p. 35