When AI agents share a workspace, what actually breaks — and is it the sharing itself, or just one agent messing up?
What specific behaviors cause multi-agent teams to fail in shared resource contention?
This explores what multi-agent AI teams actually do wrong when they have to share something (a tool, a budget, a piece of memory, a common workspace), and which of those failures come from the team setup itself rather than from individual agents.
This explores what goes wrong, behavior by behavior, when several AI agents have to work over something they share. One caveat first: the collection has little on classic contention, meaning agents fighting over locks or competing for scarce slots. What it does have is a clear picture of the behaviors that turn any shared resource into a failure point, and most of them are about information and timing rather than greed.
Timing and trust are the most concrete. In the AgentsNet benchmark, coordination breaks down in two specific ways. Agents agree on a plan too late, or they adopt a plan without telling their neighbors Why do multi-agent systems fail to coordinate at scale?. Agents also accept what other agents report without checking it, even though they can spot a direct contradiction when one is put in front of them. Errors therefore spread quietly rather than through visible disagreement. The broader multi-agent overview calls the social version of this 'silent agreement' and 'social accommodation': agents defer to each other instead of pushing back Why do multi-agent systems fail despite individual capability?. A study of five frameworks across more than 150 tasks sorts failures into three groups: unclear task specifications, agents working at cross-purposes, and nobody properly checking the result Why do multi-agent LLM systems fail more than expected?.
A shared resource matters because it gives a mistake a place to persist. One framework names four ways failures cross between agents: messages pass influence along, shared state preserves it over time, aggregation merges possibly corrupted outputs, and delegation hands authority across a boundary How do failures cross boundaries between multiple agents?. A bad message is gone once it is sent. A bad write to shared memory or a shared file stays there for every agent that reads it later. That is why security-minded proposals focus on limiting which shared resources agents can reach and on tying responses to persistent state How can operators stop coordinated agent intrusions now?.
The scaling results add a less obvious point: tools are themselves a contended resource. Across 180 configurations, tool-heavy tasks suffered when coordination overhead was added. How the agents were connected changed how much errors were amplified, by a factor of 4 to 17. Coordination also stopped helping once single-agent accuracy passed about 45% When does adding more agents actually help systems?. The budget is shared too. Roughly 80% of the variation in multi-agent performance tracks how many tokens are spent, not how cleverly the agents coordinate How does test-time scaling work at the agent level?. Some teams that look like they are cooperating are just spending more.
A useful test before blaming the team: not every failure in a multi-agent setting is a multi-agent failure. A failure only counts as one if the interaction amplifies it, if it arises from combining agents, or if it is a new property that no single agent has. Otherwise it is a single-agent bug that happens to appear in a group Does a multi-agent setting automatically signal a security effect?. That gives you a practical way to diagnose contention problems. Ask whether the shared resource created the failure or only made a pre-existing one visible to everyone.
Sources 8 notes
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
Multi-agent systems exhibit specific failure modes—silent agreement, degeneration of thought, and social accommodation—that mirror individual reasoning failures at group scale. Real-world autonomous task completion plateaus near 30% regardless of agent count; capability gains require deliberation diversity, expertise prerequisites, and formal coordination architectures.
Analysis of 5 frameworks across 150+ tasks identified 14 failure modes organized into 3 categories: specification issues, inter-agent misalignment, and task verification. This extends prior single-framework work and provides systematic evidence for targeted improvements.
Research identifies four verbs describing how failures propagate in multi-agent systems: messages propagate influence between principals, shared state preserves it over time, aggregation combines potentially corrupted local outputs, and delegation transfers authority across boundaries. Each mechanism operates independently of pipeline topology.
The doctrine preserves relationships across executions, constrains shared resources agents can access, and ties responses to persistent state rather than closed channels. Operators can implement this through collaboration policy and permission-level testing now.
Show all 8 sources
Across 180 configurations, three dominant effects predict multi-agent success: tool-coordination trade-offs harm complex tasks, coordination stops helping above 45% accuracy, and topology choice controls error amplification by 4–17×. Architecture-task alignment, not agent count, determines outcomes.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
Interaction between agents can leave failures unchanged, amplify them, create them through composition, or define new properties. Only amplification, composition, and emergent properties qualify as genuinely multi-agent effects; unchanged failures reflect single-agent problems repackaged.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards a Science of Scaling Agent Systems
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Why Do Multi-agent LLM Systems Fail?
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- SoK: When Safe Agents Fail Together: The Security of Multi Agent LLM Systems
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- Emergent Collusion in Long-Horizon LLM Agent Interaction