Adding more AI agents to a task costs real money and accuracy — so when does a team actually beat one?
What quantitative costs and failure modes emerge when coordinating multiple agents?
This explores what adding more agents to a team measurably costs (tokens, accuracy, amplified errors) and the specific, repeatable ways agent teams break down.
This explores what adding more agents to a team measurably costs, and the specific ways agent teams break. The most surprising number is that about 80% of the performance variance in multi-agent systems comes from token budget, not from how smart the coordination is (How does test-time scaling work at the agent level?). Much of what looks like a "team effect" may just be a system spending more compute. The same note points to shared-KV-cache and latent-space approaches as ways to get the gains without paying full token prices.
A study of 180 configurations puts thresholds on when coordination pays off (When does adding more agents actually help systems?). Tools and coordination compete for the same budget, so tool-heavy complex tasks get worse with more agents. Coordination stops helping once a single agent is already above roughly 45% accuracy. And topology alone changes how much errors get amplified, by anywhere from 4× to 17×. The lesson is that fit between architecture and task decides the outcome, not headcount. Real-world autonomous task completion plateaus near 30% however many agents you add (Why do multi-agent systems fail despite individual capability?). That points to a structural ceiling, not a scaling problem.
The failures themselves are surprisingly systematic. Analysis of 5 frameworks across 150+ tasks found 14 failure modes in three families: bad task specification, agents talking past each other, and weak verification of results (Why do multi-agent LLM systems fail more than expected?). Groups also reproduce individual reasoning failures at larger scale. Agents fall into silent agreement, degenerate into repeating the same thoughts, or socially accommodate each other instead of pushing back (Why do multi-agent systems fail despite individual capability?). Size makes it worse. On the AgentsNet benchmark, agents either settle on a strategy too late or adopt one without telling their neighbors. They also accept neighbors' information without checking it, even though they can spot direct conflicts (Why do multi-agent systems fail to coordinate at scale?).
One line of work asks how a failure actually crosses from one agent to another. It names four routes: messages carry influence, shared state preserves it over time, aggregation combines possibly corrupted local outputs, and delegation hands over authority (How do failures cross boundaries between multiple agents?). A companion note adds a check on whether a problem is really a multi-agent one. If interaction leaves a failure unchanged, it's a single-agent problem in a group setting. Only amplification, composition, and emergent properties count as true multi-agent effects (Does a multi-agent setting automatically signal a security effect?).
Some designs cut these costs. Scoring each agent's contribution lets a system drop its weakest members at inference time (Can multi-agent teams automatically remove their weakest members?). Passing standardized documents instead of chatting cuts noise (Does structured artifact sharing outperform conversational coordination?). A decentralized team that keeps competing hypotheses and shares its failures beat a central planner by 8.33% under matched budgets (Can decentralized teams outperform central planners in long-running science?). Teams still have a reason to exist, because some tasks need parallel work, mixed expertise, and independent verification that no single agent loop can organize (Do single agents always hit organizational limits?). So the question is whether the specific structure earns back what it costs, not whether to coordinate.
Sources 11 notes
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.
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.
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.
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.
Show all 11 sources
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.
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.
DyLAN's three-step importance scoring mechanism (propagation, aggregation, selection) quantifies individual agent contributions and automatically removes uninformative agents during inference, optimizing team composition without task-specific tuning.
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
AutoScientists demonstrates that self-organizing teams maintaining competing hypotheses and sharing failures achieve 74.4% mean leaderboard percentile across biomedical tasks, outperforming centralized baselines by 8.33% under matched experimental budgets.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
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
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- How we built our multi-agent research system
- Why Do Multi-agent LLM Systems Fail?
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- LLMs Corrupt Your Documents When You Delegate
- Self-Organizing Agent Teams Learn to Reason Together