7 papers
Multi-Agent Systems
Orchestration, collaboration, roles, and why teams of agents succeed or fail.
- Papers
- 7
- Read time
- 1.9h
- Avg length
- 16m
Why Do Multi-Agent LLM Systems Fail?
Berkeley hand-read 200+ traces from 7 multi-agent LLM frameworks, found they fail 60–86% of the time, and built MAST — a 14-mode taxonomy showing most failures come from bad *system design and coordination*, not weak models.
- Read
- 16 min
Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
Stop hand-assigning roles to your agents — give them a mission, the right ordering protocol, and a strong model, and they self-organize into better teams than anything you'd design, but only if the model is smart enough to know what it doesn't know.
- Read
- 16 min
- Source
- arXiv:2603.28990
Games Agents Play: Towards Transactional Analysis in LLM-based Multi-Agent Systems
Give each agent three competing inner voices — a Parent, an Adult, and a Child, each with its own memory store — and let a referee pick which voice speaks, and your multi-agent simulations start producing the messy, repetitive, psychologically-loaded conversations real humans actually have.
- Read
- 14 min
- Source
- arXiv:2507.21354
Latent Collaboration in Multi-Agent Systems
Instead of having LLM agents talk to each other in English, let them pass their raw internal "thoughts" (hidden states and KV-caches) directly — no decoding to text — which is faster, cheaper, and actually more accurate, with zero extra training.
- Read
- 16 min
- Source
- arXiv:2511.20639
Multi-Agent Collaboration via Evolving Orchestration
Instead of wiring your agents into a fixed graph, put a single small "puppeteer" model in charge of picking which agent speaks next at every step, and train that puppeteer with reinforcement learning to get answers that are both better and cheaper.
- Read
- 18 min
TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture
Instead of making one agent smarter, run a dozen *different* tool-using agents in parallel, let them read each other's answers and refine for a couple of rounds, then have an LLM decide when to stop and vote on the winner — and you beat every other test-time-scaling method at the same cost.
- Read
- 16 min
- Source
- arXiv:2510.01279
Unifying Language Agent Algorithms with a Graph-Based Orchestration Engine (AGORA)
Build every agent reasoning strategy — from plain Chain-of-Thought to tree search to visual search — as swappable "operators" inside one shared graph-workflow engine, then run them all through the same evaluation harness, and you discover that the simple ones usually win on both accuracy and cost.
- Read
- 18 min
- Source
- arXiv:2505.24354