Adding more AI agents to a system can degrade its performance by up to 70%. That is not speculation. It is the headline finding from a controlled study by Google Research and MIT that evaluated 180 agent configurations across four benchmarks and three LLM families (OpenAI, Google, Anthropic). The same centralized architecture that boosted financial reasoning accuracy by 81% over a single agent made sequential planning tasks dramatically worse. The difference is not the agents. It is the task.
This matters because the default assumption in 2026, reinforced by framework marketing and conference demos, is that multi-agent systems are strictly better. They are not. The research gives us the first quantitative framework for predicting when adding agents helps and when it backfires.
What Google and MIT Actually Tested
The paper, titled “Towards a Science of Scaling Agent Systems,” evaluated five canonical agent architectures: single-agent, independent multi-agent, centralized (one orchestrator coordinating workers), decentralized (agents messaging peers), and hybrid (mix of centralized and decentralized). Each architecture was instantiated with models from OpenAI, Google, and Anthropic, across four agentic benchmarks.
The benchmarks covered deliberately different domains. Finance-Agent tested financial reasoning where multiple data sources could be analyzed in parallel. BrowseComp-Plus tested web navigation. PlanCraft tested sequential game planning where each step depends on the previous one. Workbench tested general tool-use workflows.
This diversity is what makes the findings credible. The researchers did not cherry-pick tasks where multi-agent systems shine. They tested across the spectrum and found that the results flip depending on task structure.
The 180-Configuration Grid
Five architectures, three LLM families, four benchmarks, multiple configuration parameters per architecture. The grid totaled 180 distinct configurations, each evaluated under controlled conditions. This is not a blog post claiming “we tried CrewAI and it worked great.” It is a systematic evaluation with statistical controls.
The resulting dataset let the researchers build a predictive model that correctly identifies the optimal coordination strategy for 87% of unseen tasks. That predictive accuracy is the real contribution: not just “what worked” but a framework for predicting what will work for your specific task.
When Multi-Agent Systems Outperform
On parallelizable tasks, the results were unambiguous. Centralized multi-agent coordination improved performance by 80.9% on financial reasoning compared to a single agent working alone.
The mechanism is straightforward. Financial analysis requires pulling data from multiple sources: earnings reports, market feeds, regulatory filings, news sentiment. A single agent has to do this sequentially, burning context window space and inference time on each retrieval before it can start reasoning. A centralized multi-agent system assigns each data source to a different agent, collects their outputs, and feeds the consolidated data to a synthesizer agent.
Why Parallelism Is the Key Variable
The performance gain is not about “more brains.” It is about decomposition. When a task breaks cleanly into independent sub-tasks that do not share state or require sequential reasoning, multiple agents reduce wall-clock time and let each agent focus its context window on a smaller problem.
This works in practice for:
- Research aggregation: Each agent queries a different source (patent databases, academic papers, market reports) and a coordinator merges the findings
- Multi-market analysis: Separate agents analyze different geographic markets or product categories simultaneously
- Code review at scale: Different agents check different aspects (security, performance, style) of the same codebase in parallel
The shared characteristic is that each agent’s work is independent until the final merge step. No agent needs another agent’s intermediate results to do its job.
When More Agents Make Things Worse
On sequential reasoning tasks, every multi-agent variant the researchers tested degraded performance. PlanCraft, which requires step-by-step game planning where each decision depends on the previous one, saw accuracy drops between 39% and 70% across all multi-agent configurations.
The reason is what the researchers call “cognitive budget fragmentation.” A sequential task requires holding a long chain of reasoning in working memory. When you split this across agents, each agent only sees its piece. The communication overhead of passing context between agents consumes tokens that would otherwise go toward actual reasoning. The agents spend their “cognitive budget” on coordination instead of thinking.
Error Propagation: The 17x Amplification Problem
Independent agents (no central coordinator) can amplify errors up to 17 times. If Agent A makes a mistake and passes it to Agent B, B has no mechanism to catch or correct it. B builds on the error, potentially compounding it, and passes the result downstream. By the time the output reaches the end of the chain, one small initial error has cascaded into a fundamentally wrong answer.
Centralized orchestration limits this to roughly 4.4x error amplification. The orchestrator validates outputs before passing them along, catching some errors before they propagate. That is still a 4.4x amplification factor, but it is a massive improvement over 17x.
This has direct implications for production systems. If you are building a multi-agent pipeline for anything with real consequences (financial decisions, medical triage, legal document generation), independent agent architectures without centralized validation are a liability.
Three Scaling Principles Every Builder Should Know
The research distills into three scaling effects that dominate multi-agent system behavior. Understanding these effects lets you predict whether a multi-agent approach will help or hurt before you build anything.
1. The Tool-Coordination Trade-off
As tasks require more tool usage (API calls, web browsing, database queries), coordination costs increase disproportionately. Each tool call is a potential failure point, and coordinating tool usage across agents adds synchronization overhead. For tool-heavy tasks, a single agent with well-designed tool access often outperforms a team of agents tripping over each other’s API calls.
The practical threshold: if your task requires more than 5 to 7 distinct tool interactions in sequence, a single agent with good tool management will likely beat a multi-agent setup.
2. Capability Saturation
Multi-agent coordination yields diminishing or negative returns once single-agent baselines exceed approximately 45% accuracy. If a single agent already handles a task reasonably well, adding more agents does not push performance much higher. The coordination overhead eats into the marginal gains.
This is counterintuitive. Teams often reach for multi-agent architectures specifically for their hardest problems. But if the hard problem is that the underlying model is not capable enough, distributing the work across multiple instances of the same model does not fix the capability gap. You just get the same wrong answer from more directions, with extra latency and cost.
3. Topology-Dependent Error Amplification
The choice of coordination topology (centralized, decentralized, independent) determines how errors propagate through the system. Centralized orchestration acts as a chokepoint that catches errors, but also as a bottleneck that limits throughput. Decentralized coordination avoids the bottleneck but lets errors flow unchecked between peers.
The research found that web navigation tasks performed better with decentralized coordination, while financial reasoning tasks performed better with centralized orchestration. There is no universal best topology. The right choice depends on whether your task prioritizes throughput (decentralized) or accuracy (centralized).
How to Pick the Right Agent Architecture
The predictive model from the research achieves 87% accuracy on unseen tasks, but you do not need a formal model to apply the core logic. Four questions get you most of the way there.
Can the task be decomposed into independent sub-tasks? If yes, multi-agent coordination is likely beneficial. If the task is inherently sequential (each step depends on the previous one), a single agent will almost certainly outperform.
Does a single agent already solve this at >45% accuracy? If yes, adding agents will yield diminishing returns. Focus on improving the single agent’s tools, prompts, or model instead.
How many tool interactions does the task require? High tool count (7+) with sequential dependencies favors a single agent. High tool count with independent tools favors parallel multi-agent.
What is the cost of errors? If errors propagate dangerously (financial, medical, legal), use centralized orchestration. If errors are cheap to fix or self-correcting, decentralized or independent architectures are fine.
The Cost Reality
Multi-agent systems increase token costs by 2x to 6x compared to single-agent approaches. Every coordination message, every context handoff, every validation check burns tokens. For a task where a single agent spends $0.10 in API costs, a multi-agent setup might spend $0.20 to $0.60 for the same task, and it might produce worse results if the task is sequential.
Before reaching for a multi-agent framework, calculate whether the task’s parallelism justifies the cost multiplier. If you are building a batch processing pipeline that runs thousands of times per day, a 5x cost increase adds up fast.
Frequently Asked Questions
When should you use multi-agent AI systems instead of a single agent?
Multi-agent systems outperform single agents on parallelizable tasks where independent sub-tasks can run simultaneously. Google Research found an 81% performance improvement on financial reasoning tasks using centralized multi-agent coordination. However, sequential tasks where each step depends on the previous one perform 39-70% worse with multi-agent architectures. The key question is whether your task decomposes into independent pieces.
How much do multi-agent AI systems cost compared to single agents?
Multi-agent systems increase token costs by 2x to 6x compared to single-agent approaches. The overhead comes from coordination messages, context handoffs, and validation checks between agents. For high-volume batch processing, this multiplier adds up significantly. Whether the cost is justified depends entirely on whether the task benefits from parallelism.
What is error propagation in multi-agent systems?
Error propagation occurs when one agent’s mistake cascades through downstream agents. Google and MIT research found that independent agents (no central coordinator) can amplify errors up to 17 times, while centralized orchestration limits amplification to roughly 4.4 times. For production systems handling financial, medical, or legal tasks, this makes centralized validation essential.
What is capability saturation in AI agent scaling?
Capability saturation means that adding agents yields diminishing or negative returns once a single agent already achieves around 45% accuracy on a task. If the underlying model handles the task reasonably well on its own, distributing work across multiple instances of that model does not fix capability gaps. It just adds coordination overhead.
How do you choose between centralized and decentralized agent architectures?
Centralized architectures route all communication through an orchestrator that validates outputs, making them ideal for accuracy-critical tasks like financial analysis. Decentralized architectures allow direct peer-to-peer agent communication, better for throughput-critical tasks like web navigation. The Google and MIT study found the optimal topology depends entirely on the specific task, with their predictive model correctly identifying the best architecture for 87% of unseen tasks.