Every agent framework pitch deck has the same slide. Specialized agents collaborate. One plans, one codes, one reviews. Emergent intelligence from the mesh. Ship faster, think deeper, scale wider.
The research says otherwise.
The numbers nobody puts on the slide
Berkeley researchers analyzed 7 popular multi-agent frameworks across 200+ tasks. Six expert human annotators. Over 15,000 lines of conversation traces per task. The results:
ChatDev, a state-of-the-art multi-agent coding framework, had correctness as low as 25%.
They found 14 distinct failure modes. Not edge cases. Structural problems that get worse as you add agents.
A separate study from Google Research and MIT Media Lab tested sequential reasoning tasks across 180 agent configurations. On PlanCraft, every multi-agent variant degraded performance by 39-70% compared to a single agent: centralized -50.4%, decentralized -41.4%, hybrid -39.0%, independent -70.0%.
A third study from Stanford showed that when you equalize thinking-token budgets, single agents match or outperform multi-agent systems on multi-hop reasoning. The MAS "gains" in benchmarks come from spending more tokens, not from smarter coordination.
The 14 ways agent meshes fail
The Berkeley taxonomy (MAST) organizes failures into three categories:
Specification and system design failures. Agents disobey task specifications. They disobey role specifications. They repeat steps. They lose conversation history. They don't know when to stop.
Inter-agent misalignment. Conversations reset unexpectedly. Agents fail to ask for clarification. Tasks derail. Agents withhold information from each other. They ignore other agents' input. Their reasoning doesn't match their actions.
Task verification and termination. Agents terminate prematurely. Verification is incomplete or incorrect.
The distribution is roughly even across categories. No single failure type dominates. This means you can't fix agent meshes by solving one problem. The failure surface is the architecture itself.
Why coordination costs more than it saves
Every agent-to-agent handoff is a lossy translation. Agent A's output becomes Agent B's prompt. Context degrades at each hop. With 4 agents in a chain, you've lost more information to serialization than you gained from specialization.
The Berkeley paper points to organizational theory for the explanation. They reference High-Reliability Organizations research from Roberts and Rousseau (1989): even organizations of sophisticated individuals fail catastrophically if the organization structure is flawed.
The failure modes they found in agent meshes directly violate the defining characteristics of high-reliability organizations. Agents overstep their roles (violating hierarchical differentiation). Agents fail to seek clarification (violating deference to expertise). These are coordination failures, not LLM limitations.
The researchers tried to fix this with better prompts and redesigned agent topologies. The result: +14% improvement for ChatDev. Still nowhere near production-ready. Their conclusion: these failures require structural redesigns, not prompt engineering.
The one exception that proves the rule
Multi-agent coding systems hit 72.2% on SWE-bench Verified versus 65% for single agents using the same model. That's real.
But look at what's actually happening. One agent generates code. Another reviews it. A third fixes the issues. This isn't a mesh. It's a pipeline. Generate, review, fix. Three steps, clear handoffs, structured output at each stage.
The adversarial pattern works: one agent creates, another critiques. The collaboration pattern doesn't: agents discussing, negotiating, building consensus.
The difference matters. A pipeline has defined interfaces between stages. A mesh has N-squared communication paths. Pipelines fail linearly. Meshes fail combinatorially.
What actually ships
The pattern that works in production is boring:
One capable agent. Good tools. Curated context. Human oversight.
I run a single CLI agent instance with file tools, shell access, and a set of steering files that took an afternoon to write. It handles daily vault triage, processes captures, manages infrastructure health checks, and generates contextual summaries. All via cron. No mesh. No orchestration framework.
Here's what a single-agent setup looks like in practice:
# Single agent. One model, good tools, curated context.
# (Strands Agents SDK / Amazon Bedrock AgentCore)
from strands import Agent
from strands.models.bedrock import BedrockModel
model = BedrockModel(model_id="eu.anthropic.claude-sonnet-4-20250514-v1:0")
agent = Agent(
model=model,
tools=[file_read, file_write, shell, web_search],
system_prompt=open("steering.md").read(),
)
result = agent("Analyze deployment logs and summarize failures")
# Total: 1 LLM call, 1 context window, zero coordination overhead.
Now the multi-agent version of the same task — an "SRE team" that teams actually try to build:
# Multi-agent. Same model split into an "SRE team."
log_parser = Agent(model=model, system_prompt="You parse logs. Extract error patterns and sequences.")
dependency_mapper = Agent(model=model, system_prompt="You trace causal chains between services.")
root_cause_analyst = Agent(model=model, system_prompt="You identify the single root cause.")
remediation_advisor = Agent(model=model, system_prompt="You provide fixes with specific commands.")
parsed = log_parser("Parse these error logs...") # extracts patterns
deps = dependency_mapper(str(parsed)) # traces dependencies
rca = root_cause_analyst(f"{parsed}\n{deps}") # identifies root cause
fix = remediation_advisor(str(rca)) # suggests remediation
# 4 LLM calls, 3 handoffs, each agent re-discovering what the previous already found.
Same model. Same capabilities. 7.5x the cost, worse results. Each handoff is a lossy translation.
Real benchmark: log analysis task on Claude Sonnet 4 via Amazon Bedrock (eu-central-1)
Single agent 4-agent SRE team Overhead Time 9.4s 70.6s 7.5x Total tokens 545 7,688 14.1x Input tokens 263 3,222 12.3x Output tokens 282 4,466 15.8x Quality Correct RCA + fix Same RCA, massively verbose No improvement The single agent identified the root cause (connection pool exhaustion leading to cascading failure) in one call. The multi-agent setup spent 14x the tokens to reach the same conclusion — with the log parser already identifying the root cause in step 1, making the other three agents redundant.
Test setup: both configurations used Strands Agents with
eu.anthropic.claude-sonnet-4-20250514-v1:0via Amazon Bedrock cross-region inference. Same task prompt (6-line production error log). Single agent: one call with an SRE system prompt. Multi-agent: log_parser → dependency_mapper → root_cause_analyst → remediation_advisor, each agent's output serialized as the next agent's input. No tools, no RAG. Pure reasoning comparison. Token counts from Bedrock usage metrics.Sample of one. The cost ratios match what teams report from their own multi-agent post-mortems.
Role definition helps. Agent boundaries don't. You can give a single agent structured steps, output formats, and personal instructions. You get the same focus without the serialization loss.
The mundane things that actually improve agent performance
The Berkeley paper's failure taxonomy reads like a checklist of things you can fix without adding agents:
Clear task specifications. Most failures start with ambiguous instructions. Fix the prompt, not the architecture.
Explicit stopping conditions. Agents don't know when to stop. A max-iterations cap is not a success criterion.
Tool error messages that help LLMs recover. Stack traces don't help. A thin wrapper with "this failed because X, try Y instead" improves recovery without adding a reviewer agent.
# Bad: raw exception, LLM sees a stack trace and hallucinates a fix
def read_file(path):
return open(path).read()
# Good: actionable error, LLM recovers without a "reviewer agent"
def read_file(path):
try:
return open(path).read()
except FileNotFoundError:
return f"Error: '{path}' not found. Use list_dir() to check available files."
except PermissionError:
return f"Error: No read permission on '{path}'. Try a different path."
A lessons-learned file the engineer updates after each failure. One line per lesson. Agent reads it at task start. Humans curate better lessons than agents reflecting on traces. The engineer saw the root cause. The agent only saw the symptom.
# lessons.md (human-curated, agent-consumed)
- Never run migrations without checking current schema version first
- pytest needs --no-header flag or output parsing breaks
- API rate limit is 100/min, batch calls in groups of 50
- The staging DB connection string is in .env.staging, not .env
# Agent loads lessons at task start. 4 lines of code, no extra agent needed.
lessons = open("lessons.md").read()
agent = Agent(
system_prompt=f"{base_prompt}\n\n## Lessons from past failures:\n{lessons}"
)
Verification as a step, not an agent. Add a validation check after the task. Don't spin up a verifier agent that introduces its own failure modes.
Per-run cost visibility. Trivial math, rarely surfaced. If you can't see what a run costs, you can't optimize it.
Three of these (stopping conditions, verification, cost visibility) overlap enough that I ended up packaging the patterns. Shape is a small open-source library that wraps any tool-calling agent with phase control, transactions with automatic compensation, budget gates that change agent behavior at thresholds, and proof traces. One Python file, zero dependencies.
These are all single-agent improvements. Implement them yourself or use Shape. Either way, none of them require a mesh, and all of them move the needle more than adding agents.
When to actually use multiple agents
Three patterns have evidence behind them:
Adversarial review. One generates, one critiques. Red team / blue team. Works because the second agent's job is to find flaws, not to collaborate.
# Adversarial review: the one multi-agent pattern that works.
# Strands Agents SDK + Amazon Bedrock. Structured interface, not free-form "collaboration."
from strands import Agent
from strands.models.bedrock import BedrockModel
model = BedrockModel(model_id="eu.anthropic.claude-sonnet-4-20250514-v1:0")
generator = Agent(model=model, system_prompt="You write code. Be concise.")
reviewer = Agent(model=model, system_prompt="You find bugs. Be ruthless.")
def adversarial_pipeline(task: str, max_rounds: int = 2) -> str:
draft = generator(task)
for _ in range(max_rounds):
critique = reviewer(f"Find flaws in this output. Be specific.\n\n{draft}")
if "NO_ISSUES_FOUND" in str(critique):
break
draft = generator(f"Original task: {task}\nCritique: {critique}\nFix the issues.")
return str(draft)
This works for three reasons. Roles are clear: one creates, one destroys. The handoff is structured: critique is always text in, text out. Iteration is bounded, so it actually terminates. A mesh can loop forever.
Fan-out parallelism. Same task, many instances. Search 50 sources simultaneously. Not really a mesh, just parallel workers with a merge step.
Capability isolation. Agent A has a code interpreter. Agent B has a browser. They can't share tools. Separation is forced by the environment, not chosen for architectural elegance.
Everything else? One agent, good tools, curated context.
Workflow orchestrators are not agent meshes
Tools like n8n, LangGraph, and CrewAI sit in an interesting middle ground. They market themselves as multi-agent platforms. They're not, really. They're deterministic pipelines with LLM-powered nodes.
n8n connects Node A to Node B to Node C. Each node might call an LLM, run a tool, or transform data. The flow is defined at design time. There's no negotiation between agents. No emergent behavior. No consensus-building.
This is the pattern that works. It's the generate-review-fix pipeline, the fan-out-merge pattern, structured handoffs with defined interfaces.
The problem starts when teams use these tools to build actual agent meshes: autonomous agents that decide at runtime which other agent to call, what to pass, and when to stop. That's where the 14 failure modes kick in. That's where the 39-70% degradation shows up.
The distinction matters:
A workflow with LLM steps is software engineering. You control the flow, the interfaces, the error handling. The LLM is a function call inside a pipeline you designed.
An agent mesh is organizational design. You define roles and hope the agents figure out the coordination. The research says they don't.
n8n used well is a pipeline. n8n used to build autonomous agent swarms is the architecture diagram that looked good in the design review.
The question worth asking
If your multi-agent system performs worse than a single agent with the same token budget, what are you paying the coordination tax for?
Usually, the answer is that the architecture diagram looked better in the design review than it does in production.
References:
Cemri et al., "Why Do Multi-Agent LLM Systems Fail?" UC Berkeley, latest revision October 2025. 7 multi-agent frameworks, 200+ tasks, 14 failure modes, MAST taxonomy. (GitHub: dataset and LLM annotator)
Kim et al., "Towards a Science of Scaling Agent Systems", Google Research and MIT Media Lab, December 2025. 180 agent configurations across four benchmarks. PlanCraft (sequential reasoning) shows 39-70% degradation across all multi-agent variants. (Google Research blog)
Tran and Kiela, "Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets", Stanford, April 2026. Under matched token budgets, single agents match or beat multi-agent systems on multi-hop reasoning.
Benkovich and Valkov, "Agyn: A Multi-Agent System for Team-Based Autonomous Software Engineering", February 2026. SWE-bench Verified: 72.2% with manager, researcher, engineer, and reviewer roles. Note: Agyn is a structured pipeline with defined handoffs, not a free-form mesh.
Roberts and Rousseau, "Research in Nearly Failure-Free, High-Reliability Organizations: Having the Bubble", IEEE Transactions on Engineering Management, 36(2), 132-139, May 1989.
Shape: single-file Python library implementing the agent governance patterns referenced in this post (phases, transactions, budget gates, proof traces).
