---
kind: concept
slug: multi-agent-systems
title: Multi-agent systems
summary: >-
  Language-model agents with roles and tools split a goal and coordinate by messages, shared memory, debate, a planner or tree search. Twenty-eight card-backed papers cover software, robots, medicine, biomedical and materials work, debate, peer review, causal discovery, simulation and surveys. Gains are mostly author-reported; some are within noise.
tags: [agents, multi-agent, orchestration, debate, simulation, causal-discovery, security, survey, reinforcement-learning, robot-code-generation, multirobot-control, multirobot-planning, ai-scientist, single-cell, provenance, materials-simulation]
papers: ["1805.00881", "2303.17760", "2305.14325", "2306.03314", "2307.04738", "2308.00352", "2308.08155", "2309.10062", "2311.10537", "2406.12708", "2406.14373", "2407.15073", "2411.16723", "2412.19669", "2502.18864", "2503.10265", "2604.15034", "2505.10468", "2507.07426", "2507.14267", "2512.04388", "2608.16578", "2108.11887", "2503.23037", "2504.19678", "2506.04133", "2404.02831", "2508.02276"]
related: [llm-agent, llm-planning, task-decomposition, tool-use, retrieval-augmented-generation, self-refinement, self-consistency, chain-of-thought-prompting, hallucination-mitigation, instruction-tuning, llm-self-evaluation, llm-as-a-judge, robot-simulation, agent-based-simulation, federated-reinforcement-learning, causal-discovery, agent-memory, prompt-injection, llm-evaluation, model-routing, policy-gradient, reinforcement-learning-verifiable-rewards, model-predictive-control, actor-critic, automated-machine-learning]
status: draft
updated_at: 2026-11-01
---

## Overview

A multi-agent system here is a group of language-model agents that work toward
one goal. Each agent has a role, its own prompt and often its own tools. A
coordinator, a shared plan or a debate joins them, and their results are
combined [arxiv:2505.10468 §3.1.2].

The agents differ mainly in how they interact:

- a fixed pipeline that passes documents between roles [arxiv:2308.00352 §3.1]
  [arxiv:2308.00352 §3.2];
- dialogue between one agent per robot, ending in a plan that rules check
  [arxiv:2307.04738 §3.1] [arxiv:2307.04738 §3.2];
- a coder and a reviewer in a group chat, optionally behind a planner, that
  write robot code from a spoken request [arxiv:2411.16723 §3.1];
- a panel of expert agents that votes and revises a shared report
  [arxiv:2311.10537 §2.4];
- a panel of expert agents that debates a model design under a critic, then
  a coding agent writes, runs and repairs the training code
  [arxiv:2508.02276 § Discussion protocol] [arxiv:2508.02276 § Code generation and repair];
- reviewer, author and area-chair agents that run a simulated review, each
  reviewer with written traits [arxiv:2406.12708 §2.1] [arxiv:2406.12708 §2.2];
- an orchestrator that routes questions to specialists and checks their
  answers [arxiv:2503.10265 §3.2] [arxiv:2503.10265 §3.3];
- a supervisor that generates, ranks and evolves ideas [arxiv:2502.18864 p. 29];
- a planner over specialists whose prompts, tools and code are edited during
  the run [arxiv:2604.15034 §4] [arxiv:2604.15034 §3.2];
- a 7B planner, trained with reinforcement learning, that writes a short plan
  for a pool of worker models: which worker runs each step, the subtask text,
  and which earlier outputs each step may read [arxiv:2512.04388 §3.1]
  [arxiv:2512.04388 app. E];
- a tree search whose moves are agent actions: retrieve, analyse, filter,
  analyse the candidates, select, or stop. Rollouts are scored and their values
  are backed up the tree [arxiv:2507.07426 §2.1] [arxiv:2507.07426 §2.2];
- a population of persona agents that vote and message over signed ties. The
  group is studied as a dynamical system, not as a team that solves a task
  [arxiv:2608.16578 §2] [arxiv:2608.16578 §3.2];
- two chat agents, an assistant and a user, that take turns under fixed
  system prompts after a task-specifier agent makes the idea concrete
  [arxiv:2303.17760 §3.1] [arxiv:2303.17760 §3.2];
- copies of one model that answer, read each other's answers and revise over
  a few rounds [arxiv:2305.14325 §2.1];
- a framework in which agents exchange messages, and a group manager or a
  program decides who speaks next [arxiv:2308.08155 §2.2];
- two debates in sequence: one picks a statistical causal-discovery algorithm
  and runs it, and a second refines the graph with text and metadata
  [arxiv:2407.15073 §3.4] [arxiv:2407.15073 §3.5];
- a society of agents with traits and memory that farm, rob, trade or donate,
  and answer each other's requests [arxiv:2406.14373 §3.3] [arxiv:2406.14373 §4];
- a fixed sequence of prompted stages for a robot team, with no messages
  between agents [arxiv:2309.10062 §4];
- a supervisor and workers that share a provenance store, where each tool
  call is checked before it runs [arxiv:2507.14267 § Methods: Safety guard].

A conceptual framework describes agents as tuples of model, role, state and
permissions, with a stateless oracle and a supervisor that halts loops
[arxiv:2306.03314 §2.1] [arxiv:2306.03314 §3.4] [arxiv:2306.03314 §3.5].

The review draws one line that the cards use. A single **AI Agent** is a
language model with tools that handles one bounded task. **Agentic AI** adds
specialized agents, planning, persistent memory and an orchestration layer
[arxiv:2505.10468 §3.1.1] [arxiv:2505.10468 §3.1.2]. The authors call this a
convention, not a formal definition [arxiv:2505.10468 §3.0.1].

The evidence is thin and mostly author-reported. One source is a narrative
review with no experiments [arxiv:2505.10468 §1.1]. One is a conceptual
framework with no experiments [arxiv:2306.03314 §6]. Another is a perspective
on biomedical agents that reports no experiments at all
[arxiv:2404.02831 §1]. The others report
experiments, several in single runs or on small samples
[arxiv:2307.04738 tab. 2] [arxiv:2502.18864 p. 11] [arxiv:2604.15034 tab. 1].
The peer-review simulation compares each setting with its own baseline, not
with real review outcomes. Its authors call that comparison unreliable,
because reviewer traits are unknown [arxiv:2406.12708 §2.4].
The debate paper uses 100-item sets for its arithmetic and GSM8K results
[arxiv:2305.14325 app. A.2]. The framework paper's group-chat comparison is
12 tasks with one run per setting [arxiv:2308.08155 app. A5]. The drone survey
reports no experiments at all [arxiv:1805.00881 §I]. The robot-code study has
63 scored samples and reports no statistical test in its text
[arxiv:2411.16723 §4]. The distributed MPC paper's timings come from one
machine [arxiv:2412.19669 tab. III]. The three newest surveys
report none either [arxiv:2503.23037 §1.2] [arxiv:2504.19678 §III]
[arxiv:2506.04133 §1]. The causal-discovery
paper's ten-seed tables overlap across methods: on Auto with Gemini, its F1 is
0.61 ± 0.08 and the LLM-SCD baseline's is 0.58 ± 0.07 [arxiv:2407.15073 tab. 2].
The simulation's 85-run means come from a proprietary model whose snapshot and
temperature are not stated [arxiv:2406.14373 §4.1] [arxiv:2406.14373 §7.2].
The planner paper's baselines are each model's best score across the authors'
own runs, private code and online leaderboards, so its comparison is not
controlled [arxiv:2512.04388 §4.2].

## Variants

### Agent tuple with creation and halting rights (conceptual)

Each agent holds a model, a role, a state, a flag for whether it may create
agents, and the set of agents it may halt. Plugins supply tools. The paper
gives no algorithm and no numbers, and it describes existing systems in these
terms [arxiv:2306.03314 §2.1] [arxiv:2306.03314 §4]. Card:
[Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents](/papers/2306.03314).

### Software roles with a shared message pool

Five roles, product manager, architect, project manager, engineer and QA
engineer, follow a fixed procedure and pass structured documents, not chat.
Each role reads only the messages it subscribed to. The engineer writes and
runs tests and debugs on failure [arxiv:2308.00352 §3.1] [arxiv:2308.00352 §3.2]
[arxiv:2308.00352 §3.3]. Author-reported: 85.9% HumanEval and 87.7% MBPP
Pass@1 [arxiv:2308.00352 §4.2]. Card:
[MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework](/papers/2308.00352).

### Per-robot dialogue with ordered plan checks

Each robot has its own LLM agent. The agents speak in turns and end with one
sub-task per robot. Ordered checks (task rules, reach, inverse kinematics,
collisions, waypoint spacing) return failures as text, and the dialogue starts
again. A centralized multi-arm planner then produces the trajectories
[arxiv:2307.04738 §3.1] [arxiv:2307.04738 §3.2] [arxiv:2307.04738 §3.3].
Zero-shot with GPT-4 on six simulated tasks, author-reported success is 0.44 to
0.95 per task [arxiv:2307.04738 tab. 2]. Card:
[RoCo: Dialectic Multi-Robot Collaboration with Large Language Models](/papers/2307.04738).

### Coder and reviewer in a group chat (robot code from speech)

A coder agent writes Python for a legged robot from a spoken request. In the
two-agent set, a reviewer checks the code for correctness, task fit and
safety, and sends it back with feedback or passes it on. A group-chat manager
picks the next speaker. The three-agent set puts a planner first, which
writes a plain-language plan for the coder [arxiv:2411.16723 §3.1]. Seven
prompts, three runs each, one observer per run [arxiv:2411.16723 §3]. The
two-agent set has the fewest first-attempt code errors: 9 points below the
single agent and 14 below the three-agent set. The three-agent set is about 5
points above the single agent [arxiv:2411.16723 §4] [arxiv:2411.16723 fig. 6].
Observer scores show no significant difference across configurations
[arxiv:2411.16723 §4] [arxiv:2411.16723 fig. 8]. Card:
[Two Heads Are Better Than One: Collaborative LLM Embodied Agents for Human-Robot Interaction](/papers/2411.16723).

### Expert panel: analyse, summarise, vote, revise

Question experts and option experts each write an analysis. A summary report
goes to a yes-or-no vote. Any objection brings a revision, for up to a fixed
number of rounds. A last prompt picks one answer. No examples are given in
any prompt [arxiv:2311.10537 §2.1] [arxiv:2311.10537 §2.2]
[arxiv:2311.10537 §2.3] [arxiv:2311.10537 §2.4] [arxiv:2311.10537 §2.5].
Author-reported average over nine medical datasets: 86.7 for GPT-4 and 72.1 for
GPT-3.5 [arxiv:2311.10537 tab. 2]. Card:
[MedAgents: Large Language Models as Collaborators for Zero-shot Medical Reasoning](/papers/2311.10537).

### Orchestrator, specialists and a rule-and-rating panel

A top-level orchestrator sends visual questions and knowledge questions to
different agents. Visual action answers pass a panel that checks a
knowledge-graph table of instrument-action pairs and two 1-to-5 ratings, with
up to three reruns. Knowledge questions get retrieved text first
[arxiv:2503.10265 §3.2] [arxiv:2503.10265 §3.3]. Zero-shot on a surgical video
benchmark, GPT-4o reaches 61.73% overall with the workflow and 31.44% without
it [arxiv:2503.10265 tab. 3]. Card:
[SurgRAW: Multi-Agent Workflow with Chain of Thought Reasoning for Robotic Surgical Video Analysis](/papers/2503.10265).

### Expert panel debates a model design, then code is written and repaired (AI scientist)

The experts share one LLM backend and differ only by system prompt. They
cover data modelling, single-cell biology and deep learning, and a critic
moderates them. Each round, the experts write plans, peers score each other's
plans, the critic writes a meta-review with a score, and the plans are
revised. Each expert's confidence is a weighted sum of its previous value, the
critic's score and the mean peer score. The loop stops when every confidence is
at least 0.8 and the spread is under 0.03, or after a round limit of 20. The
critic then picks one plan [arxiv:2508.02276 § Discussion protocol]
[arxiv:2508.02276 eq. 1]. A coding agent writes training code from that plan,
runs it, reads the errors and patches the code, for up to five attempts
[arxiv:2508.02276 § Code generation and repair]. The paper says the confidence
is not a calibrated probability or a performance measure [arxiv:2508.02276 §
Critic-mediated expert panel]. Author-reported on six perturbation datasets,
averaged over three runs: PCC 0.9883 for CPA-X on Adamson CRISPRi, and MSE_DE
0.1736 for scGen-X on Norman CRISPRa [arxiv:2508.02276 § Discussion]
[arxiv:2508.02276 § Overview]. Most per-dataset baselines appear only in
figures [arxiv:2508.02276 fig. 3] [arxiv:2508.02276 fig. 4]. Card:
[AI Scientists for Building Virtual-Cell Models](/papers/2508.02276).

### Supervisor, generation, ranking and evolution

A supervisor turns a research goal into a plan and runs six agent types from a
task queue: generation, reflection, ranking, proximity, evolution and
meta-review. Ranking is an Elo-style tournament of hypotheses. Meta-review
feedback is added to every agent's prompt, so the system changes without
fine-tuning [arxiv:2502.18864 p. 29] [arxiv:2502.18864 p. 31]
[arxiv:2502.18864 p. 34] [arxiv:2502.18864 p. 35]. On 11 goals, expert
preference ranks give an average of 2.36, against 2.45 for the best baseline
[arxiv:2502.18864 p. 11]. No public code. Card:
[Accelerating scientific discovery with Co-Scientist](/papers/2502.18864).

### Planner and specialists with self-editing resources

The AGS system has a planner and five specialists: deep research, browser use,
deep analysis, tool calling and coding. Its prompts, agents, tools,
environments and memory are registered as versioned resources. A reflect,
select, improve, evaluate and commit loop proposes edits, and a candidate is
kept only if it passes assessment [arxiv:2604.15034 §3.1] [arxiv:2604.15034 §3.2]
[arxiv:2604.15034 §4] [arxiv:2604.15034 app. G]. On GAIA test, the
non-evolving configuration scores 79.07 and the evolving one 89.04
[arxiv:2604.15034 tab. 2]. These numbers measure the edits, not the team
structure. Card:
[Autogenesis: A Self-Evolving Agent Protocol](/papers/2604.15034).

### Tree search over agent actions (drug repurposing)

The search drives the run, and the agents are its moves. Five agents and an
end action make up six actions. Four of the five agents call the same LLM. The
retrieval agent calls no model [arxiv:2507.07426 §2.3]. Monte Carlo tree search
picks which move to expand next by UCT [arxiv:2507.07426 §2.2]. A rollout's
reward is the mean of a relative score, how often the model picks the same
target, and an absolute score, the share of yes answers [arxiv:2507.07426 eq. 4]
[arxiv:2507.07426 eq. 5] [arxiv:2507.07426 eq. 6]. The setup uses
Qwen2.5-7B-Instruct without fine-tuning and 12 rollouts [arxiv:2507.07426 §3.2].
Author-reported DrugBank and KIBA recall is 0.4466 and 0.4224 with the
ground-truth count, and 0.5534 and 0.4924 with that count plus three
[arxiv:2507.07426 tab. 1]. Card:
[DrugMCTS: a drug repurposing framework combining multi-agent, RAG and Monte Carlo Tree Search](/papers/2507.07426).

### Trained planner over a worker pool (learned orchestration)

The planner is a 7B model, starting from Qwen2.5-7B. It writes one to five
steps as three equal-length lists: worker ids, subtasks and access lists. An
access list says whether a step sees no earlier context or all of it
[arxiv:2512.04388 §3.1] [arxiv:2512.04388 app. E]. Training uses GRPO with a
reward of 1 for a correct final answer, 0.5 for a wrong one that parsed, and 0
for an unparsable plan [arxiv:2512.04388 §3.1]. The
paper trains for 200 iterations of 4 questions with 64 rollouts each
[arxiv:2512.04388 app. A]. Workers are named only by ordinal id, so the planner
does not see model names [arxiv:2512.04388 app. E]. Author-reported,
unconstrained, the planner beats the three listed models on all seven
benchmarks; on LiveCodeBench v6 it scores 83.93, against 82.90 for GPT 5
[arxiv:2512.04388 tab. 1]. Under constraints (4096 output tokens, minimal
reasoning), it averages 72.35 on four tasks, against 64.14 for the best single
model and 62.13 for a mixture-of-agents baseline with eight model calls
[arxiv:2512.04388 tab. 7]. A recursive variant lets the planner call itself
again, with a cap on inner calls [arxiv:2512.04388 §3.2]. On BigCodeBench it
scores 40.0 against 37.8 without recursion [arxiv:2512.04388 tab. 2]. The
recursion fine-tune reuses questions seen in training, so these results do not
show generalisation to new questions [arxiv:2512.04388 §4.4]. The paper names no
code.
Card: [Learning to Orchestrate Agents in Natural Language with the Conductor](/papers/2512.04388).

### Staged prompts for a robot team, with no dialogue (task planning)

No agents talk to each other. Four stages run in order: decompose the
instruction into sub-tasks, form a robot or a team for each sub-task, write
code that allocates the sub-tasks and runs independent ones in parallel, and
execute the code through an interpreter. Each stage is one prompt built from
Python-style worked examples, and no stage checks the one before it
[arxiv:2309.10062 §4] [arxiv:2309.10062 §4.4]. Author-reported GPT-4 success
on a 36-task AI2-THOR benchmark is 1.00 on elemental tasks and 0.62 to 0.71
on the other three categories [arxiv:2309.10062 tab. 1]. The released script
makes three model calls, not four, and accepts only three GPT model names
[arxiv:2309.10062 § Paper vs code]. Card:
[SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models](/papers/2309.10062).

### Hierarchical planner with a provenance guard (materials simulation)

A planning supervisor splits the goal into plan steps, each naming a worker.
Workers handle structures and input files, cluster jobs and convergence. All
tool outputs go into a shared store with a provenance record. Before a tool
runs, each referenced value is checked against that store, each parameter is
judged by an LLM, and some tools get a cross-parameter judge. A report judge
walks the provenance graph before a result is accepted [arxiv:2507.14267
§ Methods: Safety guard] [arxiv:2507.14267 fig. 2] [arxiv:2507.14267 fig. 5].
The paper says a warning still registers the value and that the guard fails
open on infrastructure errors [arxiv:2507.14267 § Methods: Safety guard].
Author-reported on periodic DFT: lattice constants within about 1% MAPE of
expert DFT, and a CO/Pt(111) adsorption-energy difference within 2 meV of the
expert value [arxiv:2507.14267 tab. 2] [arxiv:2507.14267 tab. 4]. Card:
[DREAMS: Density Functional Theory Based Research Engine for Agentic Materials Simulation](/papers/2507.14267).

### Agent population on a signed social graph (simulation)

Not a task team. Thirty-two persona-conditioned LLM agents vote on binary
questions over eight rounds and exchange messages over signed ties
[arxiv:2608.16578 §2] [arxiv:2608.16578 §3.2]. A logistic update with separate
concordant, discordant and unsigned couplings predicts each agent's next vote
[arxiv:2608.16578 §4] [arxiv:2608.16578 tab. 1]. Author-reported one-step
balanced accuracy on held-out questions is 75 to 86 percent, and multi-step
rollouts score lower [arxiv:2608.16578 §5.1]. The measure is agreement with
observed votes, not correctness on a task. Card:
[Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents](/papers/2608.16578).

### Agent society with traits, memory and scarce resources (simulation)

Nine `gpt-3.5-turbo` agents share food and land. Each has sampled traits
(aggressiveness, covetousness, strength), a memory of its recent actions, and
one initiative action per day: farm, rob, trade or donate [arxiv:2406.14373
§3.2] [arxiv:2406.14373 §4]. A robbery target either resists or concedes.
Conceding makes the robber a permanent superior. A contest gives a sigmoid win
probability for the strength gap, and how that becomes one outcome is not
stated [arxiv:2406.14373 §3.3] [arxiv:2406.14373 eq. 1]. Author-reported over
85 runs: robbery falls from 0.391 to 0.037 once a sovereign exists, and farming
rises from 0.477 to 0.737 [arxiv:2406.14373 §7.2]. The run stop rule and the
pooling of runs are not explained [arxiv:2406.14373 §7]. Donation was offered
and never chosen [arxiv:2406.14373 §8.6]. The paper's covetousness prior
conflicts with its prompt range, so the trait values are not settled
[arxiv:2406.14373 §3.2] [arxiv:2406.14373 app. A]. Card:
[Artificial Leviathan: Exploring Social Evolution of LLM Agents Through the Lens of Hobbesian Social Contract Theory](/papers/2406.14373).

### Review panel with trait-labelled agents (peer-review simulation)

Three reviewer agents, an author and an area chair run five phases on one
paper: reviews, rebuttal, reviewer–AC discussion, meta-review and a batch
decision. Each reviewer's traits come from prompt text. Each run changes one
trait and is compared with a run that has none [arxiv:2406.12708 §2.1]
[arxiv:2406.12708 §2.2] [arxiv:2406.12708 §2.4]. Runs use one AC and a fixed
32% acceptance rate [arxiv:2406.12708 §2.2]. Author-reported: one malicious
reviewer changes 37.1% of decisions. Table 2 gives 62.91% agreement with
baseline for that row, and the paper does not show how the 37.1% was derived
[arxiv:2406.12708 tab. 2]. The paper gives about USD 2,780 of API spend for all
experiments [arxiv:2406.12708 app. A.2]. The authors advise against LLM
reviewers in real review [arxiv:2406.12708 §2.4]. Card:
[AgentReview: Exploring Peer Review Dynamics with LLM Agents](/papers/2406.12708).

### Copies of one model that debate over rounds

N copies of one chat model answer the question. In each later round, every
copy reads the other copies' latest answers and is asked to check them and
update its own. The method needs only text prompts, so it works with any chat
model [arxiv:2305.14325 §1] [arxiv:2305.14325 §2.1]. The paper's table uses
3 agents and 2 rounds. Its code counts the first answers as round 1, so the
same setting is one debate update [arxiv:2305.14325 tab. 1]. Author-reported,
against a single agent: arithmetic 81.8 against 67.0, GSM8K 85.0 against 77.0,
and biography factuality 73.8 against 66.0 [arxiv:2305.14325 tab. 1]
[arxiv:2305.14325 tab. 2]. The paper gives no aggregation rule in the text.
Its code takes a majority vote over the final answers [arxiv:2305.14325 §2.1].
Card: [Improving Factuality and Reasoning in Language Models through Multiagent Debate](/papers/2305.14325).

### Two-agent role-play with a task specifier (data generation)

A task-specifier agent turns an idea into a concrete task. An assistant and a
user, each with a fixed system prompt, then take turns. The user gives one
instruction per turn, the assistant answers, and the user ends the chat with
a stop token. The prompts fix the roles, forbid role flips and fix the answer
format [arxiv:2303.17760 §3.1] [arxiv:2303.17760 §3.2]. Author-reported: CAMEL
solutions win 76.3% of human votes against single-shot gpt-3.5-turbo on AI
Society, and a LLaMA-7B tuned on the generated data scores 14.0 pass@1 on
HumanEval, against 10.5 for the base model [arxiv:2303.17760 tab. 1]
[arxiv:2303.17760 tab. 3]. The paper caps conversations at 40 messages
[arxiv:2303.17760 §4]. Card: [CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society](/papers/2303.17760).

### Conversation programming with auto-reply agents (framework)

Agents are conversable: each sends and receives messages, and each can be
backed by an LLM, a human, a tool or a mix. A conversation runs by auto-reply
until a termination condition holds. Control can sit in Python, in a system
prompt, or in both [arxiv:2308.08155 §2.1] [arxiv:2308.08155 §2.2]. Built-in
patterns include a two-agent loop and a group chat run by a manager that picks
the next speaker [arxiv:2308.08155 §2.2] [arxiv:2308.08155 §3 A5]. Author-reported:
MATH full test set 69.48% against 55.18% for vanilla GPT-4 [arxiv:2308.08155 app. A1],
and ALFWorld success of 69 for a three-agent design against 54 for ReAct
[arxiv:2308.08155 app. A3]. The chat-loop limits in the library, such as 100
auto-replies and 10 group rounds, are not stated in the paper. They come from a
later library release, read through summaries, and are unconfirmed line by
line (see the card's Paper vs code section). Card:
[AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation](/papers/2308.08155).

### Debate on an algorithm, then debate on the graph (causal discovery)

Both stages use one three-role debate. An affirmative agent proposes, a
negative agent proposes a plausible alternative, and a judge picks the winner
[arxiv:2407.15073 §3.3] [arxiv:2407.15073 alg. 1]. Stage one has the debate
choose a statistical causal-discovery algorithm from a pool, and a coding agent
runs it on the full data to give a graph [arxiv:2407.15073 §3.4]. The pool is
only partly named (PC, FCI and GES) [arxiv:2407.15073 app. A.2]. Stage two turns
the graph into sentences, adds them to the metadata, and runs the debate again
on graphs [arxiv:2407.15073 §3.5] [arxiv:2407.15073 alg. 3]. Author-reported:
averaged over datasets and backbones, the method ranks first on every metric,
and with Gemini-2.0-Flash it is first on 10 of 15 evaluation points
[arxiv:2407.15073 §4.2]. The Earthquake result is matched by the Exact Search
baseline [arxiv:2407.15073 tab. 2]. Card:
[Multi-Agent Causal Discovery Using Large Language Models](/papers/2407.15073).

### Federated reinforcement learning (not an LLM team)

Listed here because its card carries this concept, but it is a different
setting. In vertical federated RL, agents may have no action or reward, and
states cannot be shared. The paper says the closest existing RL setting is
multi-agent RL, and that the two differ on these points [arxiv:2108.11887 §4.3].
Card: [Federated Reinforcement Learning: Techniques, Applications, and Open Challenges](/papers/2108.11887).
Do not transfer its results or methods to LLM agent teams.

### Decentralised patrol with a neighbour agreement (no language model)

Each drone patrols one perimeter segment. When a neighbour agrees, the two
meet at their shared border and move in the set direction. A drone turns back
at the perimeter end [arxiv:1805.00881 alg. 1] [arxiv:1805.00881 §X]. The
algorithm comes from a study the survey reviews, and the survey reports no
experiments of its own [arxiv:1805.00881 §I]. It is a multi-robot pattern, not
an LLM team, so do not transfer it to LLM agents. Card:
[Civil UAV applications survey](/papers/1805.00881).

### Distributed MPC with learned policies (no language model)

Each robot has an actor that maps its own and its neighbours' formation errors
to a control, and a critic that estimates the cost. Both are trained online in
each prediction interval, and the robot applies the actor's output
[arxiv:2412.19669 §IV-B] [arxiv:2412.19669 eq. 14]. A check against a baseline
cost resets the weights when it fails [arxiv:2412.19669 §IV-C]
[arxiv:2412.19669 eq. 18]. Author-reported, on collision avoidance with two
robots, the learned policy's per-step cost is 13.5 against 5.0 for the
IPOPT-solved baseline. Its per-step deploy time is 2e-5 s, against 0.03 to
0.05 s for the baseline [arxiv:2412.19669 tab. III]. The safe variant with
barrier terms relies on steps and proofs in an auxiliary file that is not in
the source package [arxiv:2412.19669 §V-B]. This is not a language-model team,
so do not transfer its results to LLM agents. Card:
[Toward Scalable Multirobot Control: Fast Policy Learning in Distributed MPC](/papers/2412.19669).

### Survey views: role teams, agent protocols and risk controls (no experiments)

Three surveys add no team design of their own and no measured result. The
agentic-LLM survey files role-based teams and open-ended societies under
interacting agents, and its table rows give no scores
[arxiv:2503.23037 §4.2] [arxiv:2503.23037 §4.3] [arxiv:2503.23037 tab. 3].
Card: [Agentic Large Language Models, a survey](/papers/2503.23037).

The agent review sets out a loop of planning, acting with tools and assessing
the result. The paper gives no stop rule or retry budget for it
[arxiv:2504.19678 §IV.A]. For agent-to-agent messages it describes A2A, where
agents publish capabilities as Agent Cards and tasks share an ID, and ACP,
where a session ID keeps state and retries fall to the client
[arxiv:2504.19678 §IV.C]. It calls protocol security largely unexamined
[arxiv:2504.19678 §V.F]. Card:
[From LLM Reasoning to Autonomous AI Agents: A Comprehensive Review](/papers/2504.19678).

The TRiSM review applies trust, risk and security management to LLM-based
multi-agent systems. It lists six threats, among them tool misuse, memory
contamination and multi-agent collusion or drift, and sorts risks into four
families [arxiv:2506.04133 §3] [arxiv:2506.04133 fig. 9]. Its five pillars are
explainability, ModelOps, security, privacy and lifecycle governance
[arxiv:2506.04133 §9]. Its Component Synergy Score sums, over ordered agent
pairs, the impact of one agent on another times the quality of the second
agent's output, then divides by the number of pairs [arxiv:2506.04133 §5]
[arxiv:2506.04133 eq. 4]. The paper gives one worked link (0.35 times 0.92) and
no estimator for impact or quality, and the score is not validated on real
systems [arxiv:2506.04133 §5]. Card:
[Review of trust, risk and security for LLM multi-agent systems](/papers/2506.04133).

### Collaboration schemes for biomedical discovery (perspective, no experiments)

A perspective on AI agents for biomedical research names five ways a team can
work together. In brainstorming, agents pool many ideas, and volume counts for
more than quality at first. In expert consultation, one agent or a human
gathers evidence and feeds it back to the others. In a research debate, one
team argues for a position and another critiques it. In a round-table, agents
exchange and revise proposals over several rounds. In a self-driving
laboratory, the whole discovery loop runs under a scientist's broad direction,
without step-by-step oversight [arxiv:2404.02831 §3.2] [arxiv:2404.02831 fig. 3].
It argues that heterogeneous teams, with ML tools and human experts as well as
language models, suit this work better than one model playing several roles
[arxiv:2404.02831 §3]. It names three ways to assign roles: fine-tuning,
in-context prompts and automatic role generation [arxiv:2404.02831 §3.1]. No
scheme is measured. Card:
[Empowering Biomedical Discovery with AI Agents](/papers/2404.02831).

## Single AI Agent (the contrast case)

One language model with tools, one bounded task. It runs a reason, act, observe
loop and has no coordinator [arxiv:2505.10468 §6.1]. The multi-agent designs
above add coordination on top of this loop. The clearest contrast is in the
medical cards. With GPT-4, MedAgents averages 86.7, against 80.6 for zero-shot
prompting and 85.4 for few-shot chain-of-thought with self-consistency
[arxiv:2311.10537 tab. 2]. The gain over the strongest few-shot baseline is
1.3 points.

## What the ablations show

Only some papers remove one step at a time. The rows below are the ones that
do.

| Paper and setting | Step removed or added | Measured change | Source |
|---|---|---|---|
| MedAgents, MedQA (model not stated) | Chain-of-thought, then analysis, summary and consultation | 55.0, 62.0, 65.0, 67.0 | [arxiv:2311.10537 tab. 3] |
| MedAgents, MedQA, GPT-3.5 | Six different expert domains vs six repeated domains | 64.1 vs 59.2 | [arxiv:2311.10537 tab. 7] |
| SurgRAW, GPT-4o, cognitive-inference average | Chain-of-thought prompts | 20.11 to 58.97 | [arxiv:2503.10265 tab. 4] |
| SurgRAW, GPT-4o, visual-semantic average | Panel discussion, with chain-of-thought | 46.72 to 49.81 | [arxiv:2503.10265 tab. 4] |
| RoCo, mean over six tasks | Environment feedback removed | 0.75 to 0.56 | [arxiv:2307.04738 tab. 2] |
| RoCo, mean over six tasks | Dialogue history removed | 0.75 to 0.53; higher on one task (1.00 vs 0.75) | [arxiv:2307.04738 tab. 2] |
| AgentReview, agreement with baseline decisions | Author rebuttal phase removed | 86.14% agreement (κ 0.668); agreement, not accuracy | [arxiv:2406.12708 tab. 2] |
| AgentReview, agreement with baseline decisions | Numeric ratings removed | 60.40% agreement (κ 0.052); agreement, not accuracy | [arxiv:2406.12708 tab. 2] |
| MetaGPT, HumanEval and MBPP | Executable feedback removed | Down 4.2 and 5.4 points | [arxiv:2308.00352 §4.4] |
| MetaGPT, SoftwareDev | One engineer vs four added roles | Revisions 10 to 2.5; executability 1.0 to 4.0 | [arxiv:2308.00352 tab. 3] |
| Co-Scientist, novelty rating | Search removed from reflection | Known non-novel idea rated 6.14 without search, 2.38 with | [arxiv:2502.18864 p. 38] |
| DrugMCTS, DrugBank (paper's top-k cell) | Molecule and interaction analyses both removed | 0.3472 vs 0.4466 for the full method | [arxiv:2507.07426 tab. 2] |
| DrugMCTS, DrugBank (paper's top-k cell) | Absolute reward removed, relative reward only | 0.4320 vs 0.4466 for the full method | [arxiv:2507.07426 tab. 2] |
| MAC, Survey, GPT-4o F1 | Negative agent and judge removed (single agent) | 0.36 to 0.79 | [arxiv:2407.15073 tab. 6] |
| MAC, Auto, Gemini F1 | Negative agent and judge removed (single agent) | 0.06 to 0.61 | [arxiv:2407.15073 tab. 6] |
| MAC, Cancer, GPT-4o F1 | Negative agent and judge removed (single agent) | 1.00 to 0.83, the single agent is higher | [arxiv:2407.15073 tab. 6] |
| Conductor, constrained, LiveCodeBench | Subtask text removed | 64.29 to 58.62 | [arxiv:2512.04388 tab. 9] |
| Conductor, constrained, LiveCodeBench | Few-shot examples removed | 64.29 to 54.86 | [arxiv:2512.04388 tab. 9] |
| Conductor, unconstrained, mean over AIME25, BigCodeBench, GPQA-D | GPT 5 as the only worker | 72.89 to 69.81 | [arxiv:2512.04388 tab. 10] |
| SMART-LLM, GPT-4, task scope and runs not stated | Coalition stage removed | SR 0.75 to 0.60; TCR 0.90 to 0.68 | [arxiv:2309.10062 tab. 3] |
| DREAMS, CO/Pt(111), PBE | Guard added (DREAMS to DREAMS_safe) | Steps succeeded 81% to 100%; tokens 1.8M to 23.9M | [arxiv:2507.14267 tab. 6] |

The largest measured effects in these rows come from prompting or first-stage
structure: chain-of-thought in SurgRAW, the analysis stage in MedAgents, and
feedback in MetaGPT and RoCo. The Conductor rows fit that pattern: in its
constrained setting, dropping the few-shot examples costs about 9.4 points on
LiveCodeBench, and dropping the subtask text about 5.7 (our arithmetic from
the table). The later interaction steps give mixed results. DrugMCTS is different: removing
both analysis agents costs about 0.10 DrugBank recall in the first cell, and
dropping the absolute reward costs about 0.01 [arxiv:2507.07426 tab. 2].
MedAgents' consultation adds 2 points on MedQA, and SurgRAW's panel adds 3.09
points on the visual average. RoCo's history matters on average but not on
every task. Effects of 2 to 3 points on 300 items are inside the noise: one
binomial estimate there is about 2.8 points (our arithmetic from the card)
[arxiv:2311.10537 §3.1]. In the causal-discovery paper, removing the negative
agent and the judge costs a lot of F1 on Survey and Auto, but not on Cancer with
GPT-4o [arxiv:2407.15073 tab. 6].

## When to use which

The cited results support the rows below. No paper gives a rule for how many
agents to use, and the review says so [arxiv:2505.10468 §6.1].

| Situation | Use | Why |
|---|---|---|
| One task, one capable model, latency or cost tight | One agent with chain-of-thought first; add structure only when a measurement shows a gain | The largest measured steps are prompting or analysis steps. MedAgents reports about 40 s and about $1.41 per 100 questions [arxiv:2311.10537 §3.1] |
| Subtasks need different tools or roles, and rules can check them | Coordinator, specialists and ordered checks | RoCo catches unreachable or colliding targets before motion [arxiv:2307.04738 §3.2]; SurgRAW gates answers with a tool-action table [arxiv:2503.10265 §3.3] |
| Output can be run or tested | Execution feedback loop | MetaGPT's engineer runs tests; removing feedback cost 4.2 and 5.4 points [arxiv:2308.00352 §3.3] [arxiv:2308.00352 §4.4] |
| Output is judged only by the model itself (ideas, medical answers) | Vote or tournament, and treat ratings as self-judgement | Co-Scientist's Elo comes from its own LLM judgements [arxiv:2502.18864 p. 10]. MedAgents' vote measures agreement among the same model's experts [arxiv:2311.10537 §2.4] |
| One shared scene with tight joint constraints | Try a single central planner first | In RoCo a centralized planner averages higher across six tasks: 0.81 vs 0.75 [arxiv:2307.04738 tab. 2] |
| Fixed budget, many agents | Cap rounds and calls, and count them | Autogenesis reports about 49 inference steps and 23 minutes per GAIA task, and about USD 2,000 of API spend [arxiv:2604.15034 app. B]. The round cap is a design choice [arxiv:2311.10537 §2.4] |
| Several models with different strengths, a reward you can compute on training questions, and a GPU budget for RL | A small planner trained with RL over a fixed worker pool, with workers named only by id | The Conductor beats its listed models on unconstrained tests [arxiv:2512.04388 tab. 1]. Training used 2 H100 80GB; wall-clock time is not stated [arxiv:2512.04388 app. A]. Repeat counts are unclear [arxiv:2512.04388 §4.1] |
| Worker calls must stay capped, and a learned plan can replace a fixed team | A learned plan under the same caps, compared with a fixed team and the best single model | Under the caps the planner averages 72.35 against 62.13 for a mixture-of-agents baseline with eight calls [arxiv:2512.04388 tab. 7]. Its workers get minimal reasoning, so do not read the margin as a general result [arxiv:2512.04388 app. A] |
| Robot code from spoken requests, with a person who runs and checks each output | A coder and a reviewer in a group chat, compared with one agent on the same prompts | In a seven-prompt study, the two-agent set had the fewest first-attempt errors [arxiv:2411.16723 §4] [arxiv:2411.16723 fig. 6]. Observer scores did not differ significantly [arxiv:2411.16723 §4]. The released code runs the two-agent set on a different model from the one the paper names, so treat the gap as tentative; see the [card](/papers/2411.16723) |
| Several robots with different skills, a fixed object list, and a simulator to test in | Staged prompts with few-shot examples, a coalition stage and generated code, run only in simulation | With GPT-4, SMART-LLM reaches 0.62 to 0.71 SR on the three non-elemental categories [arxiv:2309.10062 tab. 1]. Its code runs unchecked [arxiv:2309.10062 § Pitfalls] |
| Tool outputs feed a costly decision, the domain has checkable parameter rules, and a token budget of several times the unguarded run is acceptable | A supervisor with provenance-tracked tool calls, a per-call guard and a report judge | DREAMS' guarded run lifts steps succeeded from 81% to 100% on one benchmark, at about 13x the input tokens [arxiv:2507.14267 tab. 6]. The guard fails open on infrastructure errors [arxiv:2507.14267 § Methods: Safety guard] |
| No labelled data, no training budget | Prompt-only team | MedAgents and RoCo need no training [arxiv:2311.10537 §1] [arxiv:2307.04738 §1] |
| A predictive model for a dataset with no architecture chosen, a fixed held-out split, and hours of GPU time per run | An expert debate that picks the design, then code generation with a repair loop, judged on held-out data | Author-reported numbers on six datasets, over three runs each. A run takes roughly 4 to 8 GPU hours, with up to 400,000 completion tokens for a complex task [arxiv:2508.02276 tab. 2] [arxiv:2508.02276 § Discussion]. Plan scores in the paper come from LLM judges and three experts, who agree with the judges. The link to the internal confidence score is weaker [arxiv:2508.02276 § Discussion] |
| Answers can be scored, and candidates can be ranked by the model's own votes | Tree search over agent actions with a fixed rollout budget | DrugMCTS reports higher recall than single-shot GPT-4o-mini and DeepSeek-R1 on one benchmark pair [arxiv:2507.07426 tab. 1]. It keeps 12 rollouts; 8 to 12 gives clear gains, while 12 to 24 helps only one case [arxiv:2507.07426 §4.3] |
| Observed votes or messages of a group exist, and the next step is what matters | A fitted one-step dynamics rule, checked on held-out data | Fitted to observed votes [arxiv:2608.16578 §1]. One-step accuracy is 75 to 86 percent and multi-step rollouts score lower [arxiv:2608.16578 §5.1] |
| How a group's traits or roles shape a decision, with no labelled outcomes to score against | A trait-labelled panel, one trait varied at a time against a trait-free baseline; read differences as sensitivity, not accuracy | The paper uses this design and gives its baseline comparison and API cost [arxiv:2406.12708 §2.4] [arxiv:2406.12708 app. A.2]. It does not check outcomes against real reviews [arxiv:2406.12708 §2.4] |
| Causal structure from observational data plus text metadata, with several backbones available | Algorithm chosen by a debate, then a metadata debate with a judge | Best mean rank on all three metrics across five datasets [arxiv:2407.15073 §4.2]. The backbone matters: on Child, MAC's F1 is 0.44 with Gemini and 0.65 with DeepSeek-R1 [arxiv:2407.15073 tab. 2] |
| Token cost must stay low, and one model call per decision is the budget | A single debate agent with no negative agent and no judge | Competitive on most datasets at about half the cost, but much lower on Survey and Auto [arxiv:2407.15073 §4.4] [arxiv:2407.15073 tab. 6] |
| How a norm, hierarchy or resource rule might arise, studied as a model rather than as a forecast | A simulated society with traits, memory and a fixed action list; read results as this model's behaviour | Results are measured in the paper's own world, with no external dataset or baseline [arxiv:2406.14373 §5]. Model settings are not stated [arxiv:2406.14373 §4.1] |
| Many robots keep a formation, and the per-step solver is too slow for the onboard computer | A learned distributed policy with an online stability check, for the unconstrained form | Its deploy time per step is far below the solver's, but in collision avoidance the learned cost is higher at two robots [arxiv:2412.19669 tab. III]. The constraint-handling variant cannot be built from the paper alone [arxiv:2412.19669 §V-B] |
| Several RL learners that must keep data private | Federated RL, not an LLM team | The paper separates vertical federated RL from multi-agent RL [arxiv:2108.11887 §4.3] |
| A short-answer question where one model's first answer is often wrong, and several calls per question are affordable | Several copies debate for a few rounds, then take a majority or the consensus | Debate beats single answers on arithmetic and GSM8K in the paper's table [arxiv:2305.14325 tab. 1]. It can converge on a wrong answer, and the paper gives no rule for when the consensus is right [arxiv:2305.14325 §5] |
| Several chat turns must end in a named result, and a person may step in | Two-agent conversation with an end token, or a group chat with a speaker choice | The framework paper describes both patterns and a human input mode [arxiv:2308.08155 §2.2]. Its group chat was run on 12 tasks [arxiv:2308.08155 app. A5] |
| Many synthetic task conversations for training data, with no human writing each prompt | Role-play with a task specifier and fixed prompts, with a message cap | The paper reports the generated data trains a LLaMA-7B model, but cost grows with length [arxiv:2303.17760 §4]. The released model's base licence is non-commercial, so check it before use [arxiv:2303.17760 app. "Training Details"] |
| Planning how far a biomedical discovery agent should act on its own | A four-level autonomy scale (0 to 3), rated in three areas, with the lowest area setting the level | The scale is a shared vocabulary; the paper gives no test for the level boundaries and states that Levels 2 and 3 have yet to be realised [arxiv:2404.02831 §4] [arxiv:2404.02831 tab. 1]. Nothing is measured |
| Robots or drones that split an area without a language model | A neighbour-agreement rule of the kind in the drone survey | The survey gives the rule but no measurement [arxiv:1805.00881 alg. 1] [arxiv:1805.00881 §I] |
| Tool-using team that acts on real systems or regulated data | A threat review before deployment: map each threat to a control and test it | The TRiSM review lists six threats, including tool misuse and multi-agent collusion or drift [arxiv:2506.04133 §3]. It gives no measured effect for any control [arxiv:2506.04133 §6] |
| Agents built by different teams must hand tasks to each other | An agent-to-agent protocol such as A2A or ACP, with a task or session ID and client-side retry | The survey describes both protocols and their state handling, and gives no measurement [arxiv:2504.19678 §IV.C] |
| A group-level score of how much agents help each other | The synergy score as a starting point, with impact and quality defined by the implementer | Defined in eq. 4, but the paper gives no estimator and no validation [arxiv:2506.04133 §5] [arxiv:2506.04133 eq. 4] |
| Reusing code | Check the licence first | The MetaGPT and RoCo cards record MIT code, the CAMEL card Apache-2.0, and the AutoGen card MIT. The debate card records no licence, so reuse only its design. The MedAgents and SurgRAW cards record no licence file, so reuse the design only |

## Implementation sketch

The sources do not give one multi-agent algorithm. The sketch below combines
their common parts: a coordinator, specialists, ordered checks, shared memory
and a round cap. Stop rules, memory schema and checks are choices the
implementer must make.

```pseudocode
# Plan, act, check and revise in bounded rounds  [arxiv:2307.04738 §3.2]
# Input:  goal (text), roles (name -> agent), checks, max_rounds
# Output: combined answer (text), or None when no round passes

def run_team(goal, roles, checks, max_rounds):
    # Shared memory of messages and results  [arxiv:2308.00352 §3.2]
    memory = []
    # Coordinator splits the goal into one sub-task per role
    #   [arxiv:2307.04738 §3.1] [arxiv:2503.10265 §3.2]
    plan = coordinator(goal, roles, memory)
    for _ in range(max_rounds):
        # Each specialist works on its own sub-task  [arxiv:2308.00352 §3.1]
        results = {r: roles[r](plan[r], memory) for r in plan}
        memory.append(results)
        # Checks run in order. They may be rules (RoCo) or model ratings
        # (SurgRAW); the first failure becomes the feedback
        problem = first_failure(checks, results)
        if problem is None:
            return combine(goal, results)
        # Feedback goes back to the coordinator for a new plan
        plan = coordinator(goal, roles, memory + [problem])
    # Round cap reached: report failure, do not guess
    #   [arxiv:2311.10537 §2.4]
    return None
```

## Contradictions

- **Team versus one central planner.** In RoCo the centralized planner, which
  sees the full scene, averages 0.81 over six tasks. The dialogue averages
  0.75. The dialogue is higher on two tasks, Sort Cubes (0.93 vs 0.70) and Move
  Rope (0.65 vs 0.50) [arxiv:2307.04738 tab. 2]. RoCo is the only card that
  compares a team with a centralized planner. Its own text says the dialogue is
  sometimes comparable to the central planner [arxiv:2307.04738 §5.1].
- **Medical gains do not carry over in one comparison.** MedAgents on GPT-4
  averages 86.7 across nine medical question sets [arxiv:2311.10537 tab. 2].
  In the SurgRAW table, the run of MedAgents on GPT-4o scores 31.96% overall,
  and MDAgents scores 32.88%. Both sit close to zero-shot GPT-4o at 31.44%
  [arxiv:2503.10265 tab. 3]. The backbones, data and tasks differ, so these
  do not refute each other. They do show that the team design did not add to
  the single model on the surgical benchmark.
- **Paper and code disagree on basic settings.** The MedAgents paper uses
  temperature 1.0 and up to five rounds; its repository uses 0 and three
  [arxiv:2311.10537 §3.1]. MetaGPT's paper allows three retries; its
  repository allows five test rounds [arxiv:2308.00352 §3.3]. Reproduced
  numbers may differ. See the cards' Paper vs code sections:
  [/papers/2311.10537](/papers/2311.10537), [/papers/2308.00352](/papers/2308.00352).
- **Team structure is not checkable in the released code.** The Autogenesis
  paper describes a planner and specialists. Its repository's HLE config says
  the old planner pipeline is not part of the release, and the answering
  pipeline is a single agent [arxiv:2604.15034 §4]. See the
  [Autogenesis card](/papers/2604.15034).
- **Search versus team.** DrugMCTS beats single-shot models with and without
  RAG in its headline table [arxiv:2507.07426 tab. 1]. Its ablation rows do not
  include one agent run inside the same search. Its relative reward alone comes
  close to the full method [arxiv:2507.07426 tab. 2]. So the gain mixes search,
  agent steps and the scoring rule, and cannot be credited to the team alone.
- **Single agent versus debate, by dataset (causal discovery).** The text says
  the single-agent variant is "competitive" on most datasets, and that Earthquake
  and Cancer are comparable [arxiv:2407.15073 §4.4]. The table shows the single
  agent at F1 1.00 on Cancer with GPT-4o, against 0.83 for the full method
  [arxiv:2407.15073 tab. 6]. The table also shows large gaps on Survey and Auto
  [arxiv:2407.15073 tab. 6]. The text and table do not agree on Cancer.
- **Simulated message scheme.** The Physics of Agents paper says each agent
  writes one message per round [arxiv:2608.16578 §2]. Its code says two of the
  four models were generated with one message per recipient. The card's Paper
  vs code section records this. See the [card](/papers/2608.16578).
- **Review versus measured results.** The review lists collaborative medical
  decision support as an application area. It gives no measurement for it
  [arxiv:2505.10468 §4.0.2]. MedAgents reports measured gains on medical
  multiple-choice questions [arxiv:2311.10537 tab. 2]. Do not cite the review
  as evidence for the MedAgents result.
- **Text and table differ in the peer-review paper.** The AgentReview text gives
  69.8% agreement for authoritarian area chairs. Its table gives 69.41%, with
  the same Cohen's κ of 0.266 in both [arxiv:2406.12708 §3.2]
  [arxiv:2406.12708 tab. 2].
- **Debate helps in one paper and not in another.** The debate paper reports
  arithmetic accuracy of 81.8 with three agents against 67.0 for one agent
  [arxiv:2305.14325 tab. 1]. The AutoGen paper says its authors did not run
  multi-agent debate on the full MATH set, because on the smaller set it did
  worse than vanilla GPT-4 [arxiv:2308.08155 app. A1]. The models, tasks and
  prompts differ, so neither result refutes the other. Do not quote either as
  settling whether debate helps.
- **Does a team beat the best single model?** Under caps, the learned planner
  averages 72.35 against 64.14 for the best single model [arxiv:2512.04388
  tab. 7]. MedAgents adds 1.3 points over its strongest few-shot baseline on
  medical questions [arxiv:2311.10537 tab. 2], and in SurgRAW the MedAgents run
  on GPT-4o scores 31.96% against 31.44% for zero-shot GPT-4o
  [arxiv:2503.10265 tab. 3]. The planner's setting uses minimal worker
  reasoning and a 4096-token cap, which the other two papers do not share. The
  three results therefore do not settle the question between them.
- **Agents reading each other: a gain or a risk.** Copies of one model that
  read each other's answers reach 81.8 on arithmetic, against 67.0 for one
  agent [arxiv:2305.14325 tab. 1]. The same paper says the consensus can settle
  on a wrong answer [arxiv:2305.14325 §5]. The TRiSM review lists group-level
  errors, where agents reinforce each other, as a risk family
  [arxiv:2506.04133 §3] [arxiv:2506.04133 fig. 9]. The review measures
  nothing for that risk, so the two sides are not tested against each other.
- **Two agents beat three, or the gap is not significant.** The robot-code paper
  concludes that the two-agent set performs better than the three-agent one
  [arxiv:2411.16723 §5]. Its results text says observer scores do not differ
  significantly across configurations [arxiv:2411.16723 §4]. The Discussion also
  calls the effect significant, and the paper reports no test
  [arxiv:2411.16723 §5]. The card treats the conclusion as tentative.
- **Cost metrics disagree in the distributed MPC paper.** Per-step cost is lower
  for the learned policy with no obstacles [arxiv:2412.19669 tab. III]. The
  cumulative cost at two robots favours the solver, 8.9 against 16.9
  [arxiv:2412.19669 §VI], and the collision-avoidance per-step cost at two
  robots favours it too, 5.0 against 13.5 [arxiv:2412.19669 tab. III].
- **Paper and repository disagree on the design loop.** The paper sets 20
  discussion rounds and a confidence update with weights 0.3, 0.4 and 0.3
  [arxiv:2508.02276 § Discussion protocol] [arxiv:2508.02276 eq. 1]. The
  repository's pipeline runs at most six rounds and uses other weights. The
  paper names OpenHands as the code agent, and the repository defaults to
  Codex. The card's Paper vs code section lists these differences:
  [/papers/2508.02276](/papers/2508.02276). Results from the paper's design
  cannot be assumed to follow from the released code.

## Papers

- [AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges](/papers/2505.10468): narrative review that separates single-entity AI Agents from multi-agent Agentic AI, with qualitative comparisons and a list of challenges and mitigations.
- [Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents](/papers/2306.03314): conceptual framework of agent tuples, plugins, an oracle and a supervisor; no experiments.
- [MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework](/papers/2308.00352): five software roles that pass documents through a message pool, with executable feedback.
- [RoCo: Dialectic Multi-Robot Collaboration with Large Language Models](/papers/2307.04738): one LLM agent per robot in a dialogue, with ordered plan checks and a centralized motion planner.
- [MedAgents: Large Language Models as Collaborators for Zero-shot Medical Reasoning](/papers/2311.10537): role-play experts analyse, vote and revise a report for zero-shot medical multiple-choice questions.
- [SurgRAW: Multi-Agent Workflow with Chain of Thought Reasoning for Robotic Surgical Video Analysis](/papers/2503.10265): orchestrator, specialists and a rule-and-rating panel for surgical video questions.
- [Accelerating scientific discovery with Co-Scientist](/papers/2502.18864): supervisor and six agent types that generate, rank and evolve research hypotheses.
- [Autogenesis: A Self-Evolving Agent Protocol](/papers/2604.15034): registry of versioned agent resources and a commit-on-accept update loop, used with a planner and specialists.
- [Learning to Orchestrate Agents in Natural Language with the Conductor](/papers/2512.04388): a 7B planner trained with GRPO writes a short plan of worker models, subtasks and visibility lists; no code is named.
- [DrugMCTS: a drug repurposing framework combining multi-agent, RAG and Monte Carlo Tree Search](/papers/2507.07426): five LLM-backed agents as moves in a tree search for drug-target recall, without fine-tuning.
- [AgentReview: Exploring Peer Review Dynamics with LLM Agents](/papers/2406.12708): simulated five-phase review with trait-labelled reviewers, an author and an area chair; one trait varied at a time.
- [Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents](/papers/2608.16578): persona agents voting over signed ties, with a fitted one-step update; a simulation study.
- [Federated Reinforcement Learning: Techniques, Applications, and Open Challenges](/papers/2108.11887): survey of federated RL, including a vertical variant that the paper separates from multi-agent RL.
- [CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society](/papers/2303.17760): two role-playing chat agents with a task specifier and fixed prompts, used to generate training conversations.
- [Improving Factuality and Reasoning in Language Models through Multiagent Debate](/papers/2305.14325): copies of one chat model answer, read each other and revise over rounds, with measured gains on reasoning and factuality.
- [AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation](/papers/2308.08155): a framework of conversable agents, with auto-reply, a group chat manager and several example applications.
- [Civil UAV applications survey](/papers/1805.00881): drone survey with a decentralised patrol rule and no experiments; not an LLM team.
- [Artificial Leviathan: Exploring Social Evolution of LLM Agents Through the Lens of Hobbesian Social Contract Theory](/papers/2406.14373): nine chat agents in a food-and-land world, with traits, memory and robbery or contracts; a simulation with no public code.
- [Multi-Agent Causal Discovery Using Large Language Models](/papers/2407.15073): two debate stages that choose a statistical causal-discovery algorithm, then refine the graph with metadata.
- [Empowering Biomedical Discovery with AI Agents](/papers/2404.02831): perspective with four autonomy levels for biomedical agents, five collaboration schemes and a module list; no experiments.
- [Agentic Large Language Models, a survey](/papers/2503.23037): survey that sorts agentic LLMs into reasoning, acting and interacting; role-based teams and societies sit under interacting, with no experiments.
- [From LLM Reasoning to Autonomous AI Agents: A Comprehensive Review](/papers/2504.19678): survey of agent benchmarks, frameworks and the MCP, ACP and A2A protocols; no experiments.
- [Review of trust, risk and security for LLM multi-agent systems](/papers/2506.04133): review of trust, risk and security for LLM multi-agent systems, with a threat list and an unvalidated synergy score.
- [Two Heads Are Better Than One: Collaborative LLM Embodied Agents for Human-Robot Interaction](/papers/2411.16723): coder and reviewer agents, with an optional planner, write robot code from spoken requests; a small study with no significant observer difference.
- [Toward Scalable Multirobot Control: Fast Policy Learning in Distributed MPC](/papers/2412.19669): learned actor and critic per robot replace a per-step solver in distributed MPC; not a language-model team.
- [AI Scientists for Building Virtual-Cell Models](/papers/2508.02276): expert agents debate a perturbation-model design under a critic, then a coding agent writes and repairs the training code; author-reported results on six datasets.
- [SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models](/papers/2309.10062): four prompted stages turn one instruction into code for a robot team in AI2-THOR; no agent dialogue, and no retry between stages.
- [DREAMS: Density Functional Theory Based Research Engine for Agentic Materials Simulation](/papers/2507.14267): planner and DFT, HPC and convergence workers with a provenance store, a per-call guard and a report judge; author-reported DFT results.