Klarna’s single AI agent handled 2.3 million customer conversations in its first month, replaced 700 human agents, and saved the company $60 million. It did this without a multi-agent swarm, without an orchestration framework, without a supervisor agent delegating to specialist sub-agents. One model, one prompt, one task scope. And it worked until Klarna tried to stretch it beyond that scope, at which point repeat contacts jumped 25% and the company started rehiring humans.
The lesson most people took from Klarna was “AI agents work.” The lesson they should have taken: a single well-scoped agent outperforms most multi-agent architectures, and the urge to add more agents is itself the trap.
The Math That Kills Multi-Agent Systems
A Towards Data Science analysis called “The Multi-Agent Trap” quantified what practitioners had been feeling for months: adding agents to a system does not add capability linearly. It adds failure modes exponentially.
Compound Reliability Decay
If a single agent completes a step with 99% reliability, that sounds excellent. Chain ten sequential steps across a multi-agent pipeline and your end-to-end reliability drops to 90.4%. At 95% per step, you are down to 59.9% overall. At 90% per step (common for complex reasoning), ten agents give you 34.9% reliability. One in three runs fails.
This is not theoretical. The companion TDS analysis “Why Your Multi-Agent System is Failing” measured a 17.2x error amplification factor in independent multi-agent systems. Without a mechanism for agents to check each other’s work, errors do not just propagate: they compound. Centralized architectures with a supervisor agent contained amplification to 4.4x, which is better but still means every error quadruples.
The Coordination Tax
Coordination overhead accounts for 36.9% of all multi-agent system failures. Two agents receiving ambiguous instructions interpret them differently. A planner agent generates a task graph, but the executor agent parses it slightly wrong. The billing agent and the refund agent both update the same customer record with conflicting data.
There is also a hard cost. A document analysis workflow consuming 10,000 tokens with a single agent requires 35,000 tokens across a four-agent implementation: a 3.5x cost multiplier before you account for retries and error handling. For an enterprise processing 100,000 documents per month, that is the difference between a $15,000 and a $52,500 monthly API bill.
The Four-Agent Ceiling
The TDS analysis identified a coordination threshold: gains from adding agents plateau beyond four. Below four agents, specialization benefits outweigh coordination costs. Above four, the overhead of inter-agent communication, state synchronization, and conflict resolution consumes whatever capability the additional agents bring.
This matches what Google and MIT found when they tested 180 agent configurations across different task types. Their study, published in December 2025, delivered the first quantitative scaling principles for agent systems.
Google’s 180-Configuration Study: The Scientific Verdict
Google Research and MIT tested agent systems at scale to answer a question the industry had been dodging: when does adding agents actually help?
Their results split cleanly by task type.
Sequential reasoning tasks (planning, multi-step logic, debugging): every multi-agent variant degraded performance by 39-70% compared to a single agent. The reason is specific: communication overhead fragments the reasoning chain. When a single agent reasons through a problem, it maintains full context across all steps. Split that reasoning across multiple agents and each one works with a partial picture. The handoff points between agents become information bottlenecks where nuance gets lost.
Parallelizable tasks (financial analysis across multiple datasets, research across multiple sources): centralized coordination improved performance by 80.9%. When the work genuinely decomposes into independent sub-problems that can run simultaneously, multi-agent systems shine.
The study introduced a predictive model that identifies the optimal architecture for 87% of unseen tasks. The core variable: can the task be decomposed into independent sub-problems, or does it require sequential reasoning? If sequential, use one agent. If parallelizable, use multiple agents with a central coordinator.
This is not a matter of taste. It is measurable.
Klarna’s $60 Million Lesson
Klarna’s AI customer service agent became the poster child for enterprise AI in February 2024. The numbers were genuinely impressive:
- 2.3 million conversations in its first month
- Two-thirds of all customer service chats handled autonomously
- Resolution time dropped from 11 minutes to under 2 minutes
- Equivalent work of 700 full-time agents
- Available in 23 markets, 35+ languages, 24/7
The financial impact was concrete. Cost per customer service transaction dropped 40% over two years: from $0.32 in Q1 2023 to $0.19 in Q1 2025. Projected savings of $40 million turned into actual savings of $60 million.
But Klarna did not build a multi-agent swarm. They built one agent with access to one knowledge base, scoped to handle the 60-70% of tickets that follow predictable patterns: order tracking, refund policies, FAQ answers. One model. One prompt. One job.
Where the Single Agent Broke
By mid-2025, repeat contacts had jumped 25%. One in four customers came back because their issue was not actually resolved. Customer satisfaction reportedly dropped around 22%. CEO Sebastian Siemiatkowski admitted publicly: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”
Klarna started rehiring human agents and shifted to a hybrid model. The instinct of many observers was: “See, they needed a multi-agent system.” But that misreads the failure. The single agent did not break because it was single. It broke because Klarna expanded its scope beyond what one agent could handle: complex disputes, multi-system complaints, emotionally charged conversations that require judgment, not pattern matching.
The fix was not more agents. It was better scoping of what the single agent should handle and human escalation for the rest.
The Academic Evidence: 41-86% Failure Rates
A March 2025 study analyzed 1,600+ annotated traces across seven popular multi-agent frameworks and found failure rates of 41-86.7% in production. The breakdown reveals something counterintuitive: 79% of failures originate from specification and coordination issues, not from the underlying models.
The researchers introduced the MAST Framework (Misalignment, Ambiguity, Specification errors, Termination gaps) to categorize failures:
Misalignment: agents interpret the same instruction differently. Agent A thinks “summarize the document” means three bullet points. Agent B thinks it means a paragraph. The orchestrator accepts both formats and produces inconsistent output.
Ambiguity: handoff protocols leave gaps. Who handles edge cases that fall between two agents’ domains? The answer is usually nobody, which means the customer hits a dead end.
Specification errors: the prompt that works for one agent produces unexpected behavior when another agent passes it data in a slightly different format.
Termination gaps: the system does not know when to stop. An error-handling agent keeps retrying a failed operation because nobody specified a termination condition.
The maximum improvement achieved with better system design (same underlying model) was 15.6%. The architecture ceiling is real.
The Industry Debate: Cognition vs. Anthropic
The tension between single and multi-agent approaches crystallized in June 2025 when two companies published contradictory recommendations within 24 hours of each other.
Cognition Labs (the team behind Devin, the AI coding agent) published “Don’t Build Multi-Agents”. Their argument: multi-agent architectures create fragile systems because of poor context sharing and conflicting decisions. They gave a concrete example: in a Flappy Bird game build, one agent created green pipes with specific hitboxes while another created the bird asset with incompatible dimensions. Neither had context of the other’s work. The result was a game where the bird could not physically navigate the pipes.
Cognition’s conclusion: “Context engineering is effectively the #1 job of engineers building AI agents.” Improving single-agent context will be more productive than splitting work across multiple agents that each see less.
Anthropic responded with “How We Built Our Multi-Agent Research System”, showing their Claude Opus 4 lead agent with Claude Sonnet 4 sub-agents outperformed single-agent Claude Opus 4 by 90.2% on internal research benchmarks. Token usage alone explained 80% of the variance in performance.
The catch: Anthropic’s use case was specifically parallelizable research. Finding board members of all S&P 500 IT companies is a task that decomposes perfectly into 500 independent lookups. There is no sequential dependency, no shared state, no coordination beyond “collect results.”
Both companies were right. They were just solving different types of problems.
When to Use One Agent vs. Many
The data points to a decision framework that is simpler than most architects want to admit:
Use a single agent when:
- The task requires sequential reasoning (one conclusion builds on the previous)
- Full context is necessary for each step (legal analysis, debugging, document review)
- The workflow processes items one at a time
- You need reliability above 90% end-to-end
- Your budget does not accommodate 3.5x token overhead
Use multiple agents when:
- The work decomposes into genuinely independent sub-tasks
- Sub-tasks can run in parallel without sharing state
- A central coordinator can merge results without resolving conflicts
- The parallelization speed gain justifies the coordination cost
- You stay at or below four agents
Do not use multiple agents just because:
- Your workflow has multiple steps (sequential steps are better handled by one agent with a clear prompt)
- You want “specialization” (one agent with role-specific instructions and tools often outperforms two agents communicating)
- You saw a multi-agent demo that looked impressive (demos optimize for impressiveness, production optimizes for reliability)
- Your vendor sells a multi-agent platform (Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027)
McKinsey’s “One Year of Agentic AI” report put it bluntly after analyzing 50+ agentic AI builds: “It’s not about the agent; it’s about the workflow. Agents aren’t always the answer.”
Frequently Asked Questions
What is the multi-agent trap in AI?
The multi-agent trap is the tendency to add more AI agents to a system when a single well-scoped agent would perform better. Research shows that chaining agents creates compound reliability decay (99% per-step reliability drops to 90.4% across 10 agents), a 3.5x cost multiplier, and coordination overhead that accounts for 37% of all multi-agent system failures.
When should you use a single AI agent instead of a multi-agent system?
Use a single agent for sequential reasoning tasks (planning, debugging, document review) where full context is needed at every step. Google’s 180-configuration study found that multi-agent variants degraded performance by 39-70% on sequential tasks compared to single agents. Multi-agent systems only outperform when the work decomposes into genuinely independent, parallelizable sub-tasks.
How did Klarna’s single AI agent replace 700 human agents?
Klarna deployed one AI agent (powered by OpenAI) scoped to handle predictable customer service queries: order tracking, refund policies, and FAQ answers. It handled 2.3 million conversations in its first month, cut resolution time from 11 minutes to under 2 minutes, and saved $60 million. The agent worked well within its scope but struggled when Klarna expanded it to complex disputes, leading to a 25% increase in repeat contacts.
What is compound reliability decay in multi-agent AI systems?
Compound reliability decay means that each agent in a sequential chain multiplies failure probability. If each agent step has 95% reliability, ten steps produce only 59.9% end-to-end reliability. At 90% per step (common for complex reasoning), you get 34.9% reliability across ten agents. Independent multi-agent systems amplify errors by 17.2x without error-checking mechanisms between agents.
What is the optimal number of AI agents in a multi-agent system?
Research identifies a four-agent ceiling: coordination gains plateau beyond four agents. Below four, specialization benefits outweigh coordination costs. Above four, the overhead of inter-agent communication, state synchronization, and conflict resolution consumes whatever capability additional agents bring. Google’s research confirmed that centralized coordination with a small number of agents works for parallelizable tasks, while sequential tasks should use a single agent.