21 / 21 studies
2026 · arXiv versionarXiv:2605.29790v1
Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems
Meta-Team preserves execution context and uses post-task communication to improve agent behavior, coordination and team organization.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Turn retrospectives into small protocol changes to evaluate in the next version.
Limitation
A 2026 preprint; distinguish learning from execution from adaptation to test data.
Source record
Zhezheng Hao et al.
Publication: 2026-05-28
Abstract review · Checked: 2026-09-06
2025 · arXiv versionarXiv:2512.08296v3
Towards a Science of Scaling Agent Systems
Version 3 compares 260 configurations across six benchmarks; coordination benefits depend on task structure and communication overhead.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Compare single-agent, independent parallel and coordinated teams at the same aggregate budget.
Limitation
Reported counts changed between versions; this entry uses v3. Findings are not a universal scaling law.
Source record
Yubin Kim et al.
Publication: 2025-12-09
Reviewed revision: 2026-04-08
Abstract review · Checked: 2026-09-06
2025 · arXiv versionarXiv:2503.13657v3
Why Do Multi-Agent LLM Systems Fail?
Version 3 presents a dataset of over 1,600 traces and 14 failure modes grouped into design, inter-agent misalignment and verification.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Log failure causes such as wrong tasks, broken handoffs and premature termination alongside outcomes.
Limitation
The taxonomy is a substantial starting point, not a claim to cover every failure in all systems.
Source record
Mert Cemri et al.
Publication: 2025-03-17
Reviewed revision: 2025-10-26
Abstract review · Checked: 2026-09-06
2025 · arXiv versionarXiv:2510.10047v1
SwarmSys: Decentralized Swarm-Inspired Agents for Scalable and Adaptive Reasoning
SwarmSys combines explorer, worker and validator roles with adaptive matching and pheromone-inspired reinforcement.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Tie trail strength to verified task outcomes rather than an agent's self-reported confidence.
Limitation
Preprint results belong to the authors' experiments; this app has not reproduced that performance.
Source record
Ruohao Li et al.
Publication: 2025-10-11
Abstract review · Checked: 2026-09-06
2023 · arXiv versionarXiv:2308.00352v7 · ICLR 2024
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
MetaGPT combines standardized workflows and specialist roles in a software production pipeline with intermediate verification.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Tie handoffs to inspectable intermediate artifacts and contracts.
Limitation
Standardization alone is not correctness; evaluate integration and test outcomes separately.
Source record
Sirui Hong et al.
Publication: 2023-08-01
Reviewed revision: 2024-11-01
Abstract review · Checked: 2026-09-06
2024 · arXiv versionarXiv:2407.01489
Agentless: Demystifying LLM-based Software Engineering Agents
Agentless evaluates a simple approach split into localization, repair and patch validation.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Include a simple fixed workflow when evaluating a swarm.
Limitation
The 2024 benchmark and model conditions are not a current product ranking.
Source record
Chunqiu Steven Xia, Yinlin Deng, Soren Dunn, Lingming Zhang
Publication: 2024-07-01
Abstract review · Checked: 2026-09-06
2024 · arXiv versionarXiv:2406.04692 · ICLR 2025
Mixture-of-Agents Enhances Large Language Model Capabilities
A layered architecture uses previous-layer outputs to produce subsequent responses; authors report improvements on selected evaluations.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Separate diverse proposers from the role that synthesizes evidence.
Limitation
Response evaluation does not establish equivalent gains in long-horizon tool-using tasks.
Source record
Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, James Zou
Publication: 2024-06-07
Abstract review · Checked: 2026-09-06
2023 · arXiv versionarXiv:2308.08155 · COLM 2024
AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
AutoGen combines models, tools and human input in customizable agent conversation patterns.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Specify message schemas, round limits and completion independently of the framework.
Limitation
A framework's capabilities do not establish superiority of a particular swarm topology.
Source record
Qingyun Wu et al.
Publication: 2023-08-16
Abstract review · Checked: 2026-09-06
2023 · arXiv versionarXiv:2305.14325
Improving Factuality and Reasoning in Language Models through Multiagent Debate
The study examines model instances debating answers over rounds and reports gains on selected tasks.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Generate initial proposals independently, then compare rationale and counterevidence.
Limitation
Measure debate cost and shared false assumptions; do not extrapolate to every task class.
Source record
Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, Igor Mordatch
Publication: 2023-05-23
Abstract review · Checked: 2026-09-06
2021 · Journal paperScience Advances 7(28), eabg9259
Self-organization in natural swarms of Photinus carolinus synchronous fireflies
Flash bursts in natural Photinus carolinus swarms propagated in a relay-like process shaped by vegetation.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Align progress through local checkpoints without coupling the entire system to its slowest agent.
Limitation
The observed flash pattern does not provide a global clock or fault-tolerance protocol for distributed software.
Source record
Raphaël Sarfati, Julie C. Hayes, Orit Peleg
Publication: 2021-07-07
Abstract review · Checked: 2026-09-06
2014 · Journal paperScience 343(6172), 754–758
Designing Collective Behavior in a Termite-Inspired Robot Construction Team
TERMES derived local rules from a target structure and demonstrated construction by three autonomous robots coordinating through the shared environment.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Define prerequisites, ownership and completion contracts for each task.
Limitation
Robotics evidence applies under constrained construction rules; free-form code generation has no equivalent guarantee.
Source record
Justin Werfel, Kirstin Petersen, Radhika Nagpal
Publication: 2014 · exact day not verified
Abstract review · Checked: 2026-09-06
2012 · Journal paperScience 335(6064), 108–111
Stop signals provide cross inhibition in collective decision-making by honeybee swarms
Nest-site scouts directed stop signals at dancers supporting other sites; an analytic model explained resolution of deadlock between equal alternatives.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Add sourced objections and bounded decision rounds alongside proposal generation.
Limitation
Cross-inhibition does not guarantee truth; independent verification must be designed for an LLM adaptation.
Source record
Thomas D. Seeley et al.
Publication: 2012-01-06
Abstract review · Checked: 2026-09-06
2010 · Journal paperScience 327(5964), 439–442
Rules for biologically inspired adaptive network design
Physarum networks achieved comparable trade-offs in efficiency, cost and fault tolerance in a Tokyo rail-network example.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Retain communication edges for utility relative to cost, not frequency alone.
Limitation
Physarum is a multinucleate single-celled organism, not a colony of independent LLM agents.
Source record
Atsushi Tero et al.
Publication: 2010-01-22
Abstract review · Checked: 2026-09-06
2008 · Journal paperPNAS 105(4), 1232–1237 · author manuscript
Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study
Three-dimensional starling reconstructions linked interactions to about six or seven neighbours rather than a fixed metric distance.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Test a bounded set of relevant neighbours instead of all-to-all agent messaging.
Limitation
Six or seven is a field observation, not a recommended optimum for agent communication.
Source record
M. Ballerini et al.
Publication: 2008 · exact day not verified
Abstract review · Checked: 2026-09-06
2005 · Journal paperNature 433, 513–516
Effective leadership and decision-making in animal groups on the move
In a numerical model, a small informed minority could guide the group without members recognizing who was informed.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Use a specialist as a carrier of critical knowledge rather than a controller of every step.
Limitation
The evidence is a model result, not a universal specialist ratio for fish experiments or LLM teams.
Source record
Iain D. Couzin, Jens Krause, Nigel R. Franks, Simon A. Levin
Publication: 2005-02-03
Abstract review · Checked: 2026-09-06
2003 · Journal paperThe EMBO Journal 22(4), 870–881
Vibrio harveyi quorum sensing: a coincidence detector for two autoinducers controls gene expression
V. harveyi integrates signals from two autoinducer channels in a circuit regulating gene expression.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Combine multiple independent signals before triggering an expensive action.
Limitation
Chemical concentration is not a confidence score; copied agent reports are not independent evidence.
Source record
Kenny C. Mok, Ned S. Wingreen, Bonnie L. Bassler
Publication: 2003 · exact day not verified
Abstract review · Checked: 2026-09-06
1999 · Historical reviewArtificial Life 5(2), 97–116
A Brief History of Stigmergy
Stigmergy describes coordination through environmental modifications that stimulate subsequent actions.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Define the shared artifact agents read and modify before designing messaging.
Limitation
A historical review, not a performance experiment for a particular agent architecture.
Source record
Guy Theraulaz, Eric Bonabeau
Publication: 1999 · exact day not verified
Publication record review · Checked: 2026-09-06
1998 · Journal paperJAIR 9, 317–365
AntNet: Distributed Stigmergetic Control for Communications Networks
Distributed mobile agents adapt routing tables through indirect, asynchronous information exchange mediated by the network.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Update task routing from observed results rather than fixed role titles.
Limitation
Network packet costs are measurable; LLM correctness requires an additional external evaluator.
Source record
Gianni Di Caro, Marco Dorigo
Publication: 1998-12-01
Abstract review · Checked: 2026-09-06
1996 · Journal paperIEEE TSMC Part B 26(1), 29–41
Ant system: optimization by a colony of cooperating agents
Ant System combines positive feedback, distributed computation and heuristic selection for combinatorial optimization.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Reinforce useful trails, reserve exploration and decay stale trail weights.
Limitation
Requires a defined objective function; text quality is not inherently a path length.
Source record
Marco Dorigo, Vittorio Maniezzo, Alberto Colorni
Publication: 1996 · exact day not verified
Abstract review · Checked: 2026-09-06
1989 · Journal paperNaturwissenschaften 76, 579–581
Self-organized shortcuts in the Argentine ant
Argentine ants interacting through trail pheromone selected shorter routes in the experimental setup.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Record verified utility and age in shared memory, alongside the output.
Limitation
Specific species and route setup; not a universal result for ants or LLM tasks.
Source record
S. Goss, S. Aron, J. L. Deneubourg, J. M. Pasteels
Publication: 1989 · exact day not verified
Publication record review · Checked: 2026-09-06
1987 · Conference paperACM SIGGRAPH 1987
Flocks, Herds, and Schools: A Distributed Behavioral Model
Local motion rules produced collective flock animation without scripting every individual trajectory.
What can I use? Interpretation, limits and provenance
SWI engineering interpretation
Separate collision avoidance, alignment and group cohesion into distinct rules.
Limitation
A behavioral animation model, not a discovery of every interaction rule in real animals.
Source record
Craig W. Reynolds
Publication: 1987 · exact day not verified
Abstract review · Checked: 2026-09-06