Agora β€” Self-Organization in Agent Populations

From designed coordination to emergent organization.

πŸ›οΈ Agora asks: how does decentralized resource allocation shape organization in multi-agent systems? Instead of designing roles, workflows, and routing up front, Agora lets organization emerge from an economy: agents pay to work, earn by improvement, and wealth favors future activity.

Research project from the MIT–NEU–UF Summer Camp 2026.


❓ Problem

Multi-agent LLM systems coordinate well, but their organization is designed externally: fixed roles, central routing, centrally governed resources. Organization is an input, not an outcome β€” each fixed piece can become a bottleneck or a single point of failure, wasting effort and losing what works.

Prior work covers pieces of this: adaptive search allocates compute but stays globally controlled; self-evolving MAS evolve roles but treat compute and survival as exogenous; agent economies coordinate via markets but do not couple scarcity, verification, and knowledge. Agora couples all three.


🌱 How It Works

Many agents search at once Β· one official grader decides what is better Β· an economy decides who keeps working Β· what they learn is shared.

An illustrative run: identical agents with no roles submit to the grader, pay rent, earn credit along lineage, die or reproduce, and publish to a shared culture until they differentiate. Credit is the framework's settlement; culture is its public shared state.
🧩 Framework β€” the full architecture
Four planes. Discovery: agents edit and test program versions on their own, for free. Truth: one sealed grader scores submissions; improvements replace the best version, the rest become evidence. Economy: live agents pay rent, richer agents get more turns, and each gain is split across its contributors; an agent dies when its wealth runs out. Shared state: work, artifacts, skills, and evidence persist in public with causal provenance.
πŸ’° Economic selection β€” the equations

A closed loop: scarce resources create costs, verified contributions create wealth, and wealth biases who receives future compute.

  • Rent grows with resource use (model FLOPs, evaluations).
  • Wealth = starting wealth + share of each verified gain βˆ’ rent βˆ’ investment in newborns. Gains are split by causal contribution.
  • Selection: agents bid a fraction of their wealth; the next model slot goes to a wealth-weighted softmax draw.
🌱 How selection creates organization β€” static summary
Discovery β†’ Truth β†’ Causal credit β†’ Persistence β†’ Emergence. Two feedback loops drive it: economic (success changes who searches next) and cultural (success changes what others can build on). Identical agents with no roles become differentiated behavior, artifact-reuse pathways, and persistent lineages.

πŸ“Š Results

Table 1. Cross-domain frontier performance: Agora beats the target in four of five domains and ties AlphaEvolve in mathematics, for under $0.50 per task.
Task Domain Baseline Target Agora Cost Grades
PID tuning Robotics 0.0366 0.1632 Gold 0.1639 $0.43 18
MallocLab Systems 28.15 84.06 Gold 92.43 $0.46 9
Circle packing Mathematics 0.9598 2.635977 AlphaEvolve 2.635983 $0.28 12
Single-cell modality Biology 0.5467 0.7310 Gold 0.7370 $0.18 30
Holographic optics Physics 0.3927 0.8265 Gold 0.9878 $0.45 13

One run per task. Target: the task's gold reference, or AlphaEvolve's published result for circle packing. Grades: official grader calls used.

Table 2. Framework comparison on JobShop ABZ (5 instances): Agora reaches the highest score, fastest.
Framework Score ↑ Grades To best Spend
Agora 97.66 10 4.8 min $0.48
CORAL 96.57 8 6.6 min $0.57
OpenEvolve 94.83 30 8.2 min $0.31
ShinkaEvolve 94.31 29 22.8 min $0.20
AdaEvolve 91.89 30 20.9 min $0.34
EvoX 90.47 30 58.1 min $0.28

Same seed program (80.50), DeepSeek-Flash v4.1, $0.50 and 30 grades per framework, avg@3. Standard-library programs only, as the task requires. To best: start to the best grade.

Table 3. Where does organization come from? Emergent is fastest, LLM-designed is budget-efficient, human-designed is balanced.
Organization Score ↑ Grades To gold @5 min @10 min Spend Time
Emergent (E) 0.9873 17 0.7 min 0.9818 0.9873 $0.46 7.8 min
Human-designed (H) 0.9819 30 1.6 min 0.9606 0.9774 $0.42 26.6 min
LLM-designed (L) 0.9863 24 5.1 min 0.4844 0.8592 $0.29 72.6 min

Holographic optics (Frontier-Eng, physics): split a laser beam into six spots with set brightness and keep the rest dark; gold 0.8265. H: planner, reader, coder, analyst roles; L: roles the LLM writes for itself; E: no roles, selection decides. @5/@10 min: best score by then.


πŸ” Findings

Organization is an outcome of search, and its form differs across tasks:

  • Selective persistence: useful knowledge can outlive the agent that created it.
  • Recombination: winning lineages span multiple authors.
  • Culture: successful lineages promote reuse.
  • Differentiation: diverse search niches emerge from identical agents.
Designed
Self-organized
Emergent