Research focus
I’m studying how agent swarms coordinate around a shared problem: how they divide work, exchange evidence, resolve disagreement, reach consensus, and respond to conflicting views or goals. I’m especially interested in whether assigned personalities and roles change the quality of that collaboration across different kinds of tasks.
The work starts by exploring tasks with clear outcomes, then comparing team configurations under controlled conditions. Repeated runs will help separate reliable patterns from one-off conversations. As the study develops, I’ll introduce harder and longer tasks—with more constraints, dependencies, distributed evidence, and changing information—to examine when coordination helps and when it becomes overhead.
Task outcomes and collaboration process will be examined separately. Comparisons will use appropriate single-agent and independent-agent baselines, with the same model, task information, and total budget where possible.
Research questions
I’m using Chorus to study when groups of AI agents produce better work, what it costs them to get there, and what happens along the way. I’ll link each question to experiments, data, and results as the work progresses.
- Which tasks benefit from multiple agents? How do a single agent, several agents working independently, and a collaborating team compare when the work can be split up, when each step depends on the last, or when evidence is scattered across the team?
- When can a team of smaller models match a single stronger model? How do the results change when I compare one strong agent, a swarm of smaller agents, and a human working with one of those smaller agents? Does the human add the most value when planning, steering, or checking the work?
- What does it cost to reach a benchmark target? For a given score, how much money and elapsed time does each setup need, including model calls, tools, retries, and judging? How many tasks and repeat runs do I need to estimate that score reliably?
- Do assigned personas or roles improve the work? With the same model and task, how do general-purpose agents, agents with functional roles like solver or reviewer, and agents with detailed personas differ in what they contribute and how well they perform?
- What kind of diversity helps a team? Do different models, roles, or sources of evidence help agents cover each other’s blind spots? Which differences surface useful information, and which make coordination harder?
- When does disagreement improve the answer? What answer does each agent give before seeing its peers’ work, and what has it seen by the time it changes its answer? When does evidence from a correct minority change the team’s conclusion, and when does the team simply settle on the first answer it sees? If disagreement persists, when should the team keep multiple answers or seek more evidence?
- What happens when agents’ goals diverge? If agents are rewarded for different outcomes, do they still share decisive evidence and complete the shared task? How does that change as the incentives pull further apart?
Chorus
Chorus is a research harness that brings together swarms of AI agents with different personalities, roles, and perspectives to work on a shared problem statement. I created it to study how these differences shape collaboration: how agents divide work, share information, challenge each other, resolve disagreements, and reach decisions.
Through Chorus, I want to understand which tasks benefit from a team and which are better handled by one agent or several agents working independently. I’m interested in whether a swarm of smaller models can match a stronger model.
I also want to look beyond whether a team finishes a task. Does disagreement help agents discover a better answer, or do they settle too quickly on the first convincing idea? Do different personalities, roles, and sources of information help them catch each other’s mistakes? What happens when their goals conflict, or when one agent has evidence the others need?
The goal is to understand when collaboration produces better work, what makes it succeed or fail, and whether the improvement is worth the time and cost. By varying tasks and teams and repeating comparisons, I hope to build evidence about when agent swarms are useful, how they can be made more reliable, and when a simpler approach is enough.
nanoville
Nanoville remains a separate repository and connects to the Chorus engine. It is a visual workspace for running and observing benchmark sessions. It supports a scene with up to 50 participants, searchable roster, public discussion, shared canvas, and session inspection. Custom portraits and animated sprites make the team easier to follow.
The Chorus engine runs the agents and saves their work. Nanoville shows that work in an animated scene. It is not needed to run the merged Chorus app.
Repositories
- Chorus
The engine and web interface in one repository. Includes app-building experiments, smaller problems with checked answers, saved work, and one launch command.
- Nanoville
A separate visual workspace connected to Chorus, with an animated scene for following up to 50 agents, their discussion, and shared work.
Harnesses to explore
A working list of agent harnesses and collaboration systems to study alongside Chorus. The categories are rough groupings by how I plan to approach them.
Minimal implementations and tutorials
- Minimal Agent Tutorial
A step-by-step guide to building a small coding agent from scratch. Source repository.
- mini-SWE-agent
A deliberately small software-engineering agent and inspectable baseline for coding tasks.
- Swarm Kit
A compact Python multi-agent framework with handoffs, planner mode, shared state, persistence hooks, and a live dashboard.
- Miniswarm
A local-first collaboration setup where coding agents and people coordinate through a shared IRC channel.
Popular agent runtimes and orchestration
- Pi Agent Core
Pi's stateful agent runtime for tool execution, event streaming, and agent-loop control.
- Pi Durable
An experimental runtime for durable conversations, tasks, and documents designed for long-running agents.
- Gastown
A multi-agent coding workspace built around durable work items, worker agents, and coordinated delivery.
- Ruflo
A multi-agent orchestration framework for coordinating agent swarms and workflows.
Other multi-agent systems
- CAMEL
An open-source framework for building and studying communicative multi-agent systems.
- TinyTroupe
Microsoft's persona-based multi-agent simulation library for experimenting with synthetic people and environments.
- Herdr
A workspace for running and coordinating multiple coding-agent sessions across panes and machines.
Research Papers Ongoing
Currently reading
- Improving Factuality and Reasoning through Multiagent Debate
(2023) Agents propose answers, exchange them, and revise.
- Voting or Consensus? Decision-Making in Multi-Agent Debate
(2025) Compares decision protocols for multi-agent debate.
To do
- Why Do Multi-Agent LLM Systems Fail?
(2025) A taxonomy of failures in multi-agent systems.
- Capable language models can outgrow the benefits of collaboration
Yubin Kim, Ken Gu, and colleagues (2026). Compares five agent architectures across six benchmarks. See the earlier preprint, Towards a Science of Scaling Agent Systems.
- Multi-Agent Reasoning Improves Compute Efficiency: Pareto-Optimal Test-Time Scaling
Florian Valentin Wunderlich and colleagues (2026). Compares independent sampling, self-refinement, debate, and aggregation across model sizes and compute budgets.
- When combinations of humans and AI are useful: A systematic review and meta-analysis
Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone (2024). Compares human–AI teams with human-only and AI-only baselines across task types.
- AI Agents That Matter
Sayash Kapoor and colleagues (2024). Evaluates agent accuracy alongside cost and examines how agents can overfit benchmarks.
- Efficient Benchmarking of AI Agents
Franck Ndzomga (2026 preprint). Tests how well smaller task subsets recover agent rankings and absolute scores.
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
Mingqian Zheng and colleagues (2024). Tests persona prompts against a no-persona baseline on factual questions.
- Understanding Agent Scaling in LLM-Based Multi-Agent Systems via Diversity
Yingxuan Yang and colleagues (2026). Studies how differences in models, prompts, and tools affect the evidence agents contribute.
- Beyond Symmetric Agents: Cognitive Diversity and Multi-Agent Debate in Small Language Models
Leonardo Ferreira, Gardenia Liu, and Kaden Zheng (2026 preprint). Tests persona, sampling, and model diversity against independent voting at matched generation counts.
- Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?
Hyeong Kyu Choi, Xiaojin Zhu, and Sharon Li (2025). Measures the contributions of discussion and majority voting across seven benchmarks.
- Group decision making in hidden profile situations: Dissent as a facilitator for decision quality
Stefan Schulz-Hardt and colleagues (2006). Studies dissent in human groups where each person holds only part of the evidence.
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
(2023) Social interaction scenarios and evaluation for cooperating and competing agents.
- CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society
(2023) Role-playing and cooperation among language agents.
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
(2023) Evaluates model judges and documents potential biases.
- Learning Self-Orchestration in Decentralized Agent Swarms
(2026 preprint) Learned coordination and participation in decentralized swarms.
- Flocks, Herds, and Schools: A Distributed Behavioral Model
(1987) Craig Reynolds’s model of coordinated flocking behavior emerging from local rules.
- Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations
Shoham and Leyton-Brown’s book on coordination, game theory, social choice, and mechanism design.
- Kimi K2.5: Visual Agentic Intelligence
(2026) Technical report on Kimi K2.5 and its Parallel-Agent Reinforcement Learning approach to agent swarms.
- Recursive Self-Improvement through Multi-Agent Self-Supervision
(2026) Sakana’s research on searching team workflows and distilling executions into a shared model.
- AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
(2023) A framework for building applications through multi-agent conversation.
- Do as We Do, Not as You Think: The Conformity of Large Language Models
(2025) Examines conformity in language model responses.
- TinyTroupe: An LLM-powered Multiagent Persona Simulation Toolkit
(2025) A toolkit for persona-based agent simulation and behavioral studies.
- Demystifying Multi-Agent Debate: The Role of Confidence and Diversity
(2026) Studies how viewpoint diversity and confidence communication affect debate.
- Generative Agent Simulations of 1,000 People
(2024) Evaluates agents grounded in interviews with more than a thousand people.
Done
- Social Simulacra: Creating Populated Prototypes for Social Computing Systems
A prototyping method that simulates community members and their interactions to help explore how social system designs may behave at scale.
- Generative Agents: Interactive Simulacra of Human Behavior
An agent architecture that uses memories, reflection, and retrieval to plan behavior in an interactive sandbox town.
- How we built our multi-agent research system
(2025) Anthropic engineering article on a lead researcher and parallel workers.