On September 24, 2026, KPMG published its Q3 AI Pulse survey of 314 US C-suite and business leaders at organizations with $1 billion or more in revenue, fielded between July 24 and August 25. Most of the findings describe steady progress. Sixty-two percent of organizations are now building, deploying, or developing AI agents, up from 53% last quarter.
Nearly six in ten leaders report measurable business value from their AI initiatives. One number, though, moved much faster than the rest. The share of organizations actively developing or implementing multi-agent systems, in which several AI agents hand off work to one another, climbed to 25%. It had sat at 6% for the previous two quarters.
A second number moved the other way. The share of organizations building controls into their agents, alongside monitoring and evaluation, fell to 30% from 43% two quarters earlier, and KPMG notes that many are relying on guardrails they already had in place. The picture is clear: more companies are wiring agents to other agents, and fewer are designing controls into what they build.
A disclosure before we go further. Golabs designs and builds custom AI agents and the orchestration that connects them, so we benefit when companies invest in multi-agent work. KPMG sells advisory services, and one of the surveys below was commissioned by an agent platform vendor. We name each source's interest as we go, and we lean on independent academic research where it exists.
How Fast Is Multi-Agent AI Adoption Growing?
Fast. KPMG's Q3 2026 AI Pulse, a survey of large US organizations, found 25% actively developing or implementing multi-agent systems, up from 6% in each of the prior two quarters. That is roughly a fourfold rise in one quarter. Over two quarters, the share building controls into agents fell from 43% to 30%.
This is a survey of large companies. The KPMG sample is US organizations above $1 billion in revenue. We would expect mid-market adoption to be lower, but no comparable survey we found measures it. If you run a smaller company, read the 25% as a signal of where the market is heading rather than where your peers already are.
"Developing or implementing" is not "in production." The 25% includes organizations still building. That matters, because the decisions that are cheapest to change are being made right now, before most of these systems carry real workloads. In our experience, a control designed in during the build is far easier to add than one retrofitted after launch.
The jump is one quarter of data. A single survey wave can overshoot. Even so, the direction aligns with KPMG's other findings: agent work is spreading, and the share reporting significant workforce adoption of AI rose to 44%, up from 23% last quarter. Treat the exact figure with caution and the trend with respect.
Is a Team of Agents Actually Better Than One?
Not automatically, and the best independent evidence we found says the default should be skepticism. In June 2026, researchers from Salesforce Research, HKUST (Guangzhou), the University of British Columbia, and Nanyang Technological University published "The Illusion of Multi-Agent Advantage," a study of six frameworks that automatically design multi-agent systems.
They measured those frameworks against a simple baseline: one model answering several times and taking the most consistent answer. Their conclusion was that the automatically designed multi-agent systems consistently underperformed that baseline, despite being up to 10 times more expensive. In other words, the extra agents added cost without adding accuracy.
The cost gap is the striking part. On the GPQA-Diamond science benchmark using GPT-5, the single-model baseline scored 87.35% at a cost of $46.39. One automated multi-agent design scored 85.23% at $832.10. Another scored 86.75% at $636.80. Both cost more than 13 times the baseline and scored lower.
The extra agents also contributed less than their numbers suggest. In one framework, DyLAN, the agents rarely disagreed: they reached immediate, unanimous consensus in about 70% of cases with GPT-4o and more than 90% with GPT-5. And in 7 of 14 workflows produced by one framework, the "multi-agent" design turned out to be functionally identical to the simple baseline.
Two limits apply. This is a benchmark study of automatically generated systems, not of carefully engineered production workflows. And benchmarks are not your business process; they have their own data, exceptions, and approval steps. But the lesson carries over to real deployments: adding agents increases cost and does not guarantee better answers.
A second paper, "Testing Interchangeability in LLM Agent Teams," posted to arXiv on September 4, 2026, found that swapping one agent for another with the same role raised the communication a team spends per unit of progress by 16 to 63 percent, while task scores barely changed. Agents that look interchangeable on paper are not interchangeable in cost.
Our read: a multi-agent design earns its place when the work genuinely splits into parts that need different tools, data access, or permissions. Specialization can justify the overhead. Parallelism for its own sake usually cannot. We covered the bill side of this in why your AI bill is exploding.
Where Do Multi-Agent Systems Break?
At the handoffs, and in ways that are hard to trace. A survey commissioned by Kore.ai and conducted by Propeller Insights in May 2026 asked more than 400 US IT business leaders at companies with 2,000 or more employees about their AI agents. Seventy percent said they had faced a failure their teams could not trace.
Forty percent had seen a single agent failure cascade across multiple systems. Seventy-nine percent had reversed an action taken by an AI agent, and 53% said they were running agents they do not fully trust or understand. Kore.ai sells an agent platform, so read the numbers as a vendor's framing of a real problem rather than as neutral measurement.
Academic work points the same way. In "Who Broke the System? Failure Localization in LLM-Based Multi-Agent Systems," published in July 2026, researchers at the University of Louisville and the University of North Texas developed a method to identify which agent and which step caused a failure in a multi-agent system.
In their best result on one benchmark subset, using a 7-billion-parameter model, it identified the failing agent 69.05% of the time but the exact failing step only 38.10% of the time. That is a research tool under one set of conditions, not a general failure rate. It does show how hard "what went wrong?" becomes once several agents are involved.
This is the core operational difference between one agent and many. With one agent, a bad output has one author. With several, a bad output can result from a bad handoff, stale shared context, or an agent acting on another agent's mistake. If you cannot reconstruct the chain, you cannot fix it or explain it to an auditor.
Why Are Controls Falling While Agent Use Rises?
KPMG's data does not explain the drop, so what follows is our interpretation, drawn from the systems we build and review. We see three patterns repeat. Each one makes sense in the moment, and each one leaves a gap that becomes more expensive to close the longer a system runs in production.
Controls get treated as a platform feature. Teams assume the agent framework or cloud provider handles permissions and logging, so nobody inside the project owns them. KPMG's own note that many organizations rely on existing guardrails fits this pattern. Those guardrails were often designed for people and single applications, not for agents passing work to each other.
Access is copied from people. In a Gartner survey of 297 cybersecurity leaders in the second quarter of 2026, published September 30, 54% of organizations had no defined approach to limiting AI agent access, or relied on predefined human access. An agent that inherits a person's permissions and then hands off work to another agent can reach far beyond what its task requires.
Speed wins the sprint. Pilots are judged on whether they work, not on whether they can be audited, and nobody is rewarded for a permission model in a demo. The controls that would have been cheap at design time become expensive retrofits once the pilot is promoted and other systems depend on it. We covered the authentication side in securing the MCP layer.
There is also a cost to getting this wrong later. Gartner predicted in May 2026 that by 2027, 40% of enterprises will demote or decommission autonomous AI agents due to governance gaps identified only after production incidents occur. That is a prediction, not a measurement. It suggests that governance gaps found late are expensive.
For the policy layer, meaning who owns and approves agents across the company, see our AI agent governance guide. This piece is about the engineering underneath: the identities, logs, limits, and tests that let a policy actually be enforced once agents start handing work to each other.
What Controls Should a Multi-Agent System Ship With?
These are the controls we build in from the first sprint. None of them is exotic, and in our experience all of them are cheaper on day one than on day ninety. Each one answers a question that an auditor, a security lead, or an incident review will eventually ask about your system.
Justify every agent. Before adding an agent, write down what it does that one well-instructed agent could not: different data, different tools, different permissions. If the answer is "it splits the prompt," start with one agent and measure. In the benchmark study above, the simple baseline beat automatically designed multi-agent systems on both cost and accuracy.
Give each agent its own identity and the principle of least privilege. Every agent should authenticate as itself, with access scoped to its task. No agent should borrow a human's credentials or a shared service account. Separate identities also make logs meaningful: when each agent acts under its own name, you can see which one touched which record.
Make handoffs explicit and logged. Define what each agent passes to the next agent in a structured format, and record every handoff with a trace ID that follows the request end-to-end. This is what turns "something went wrong" into "step four received the wrong customer record," and it is what lets you answer an auditor.
Put humans on the irreversible steps. Payments, deletions, external messages, and access changes should require approval, no matter how confident the agents are. Seventy-nine percent of the leaders in the Kore.ai survey had already reversed an agent action. Design for that, rather than discovering it. An approval step is cheap next to an undo.
Set budgets and circuit breakers. Cap tokens, tool calls, and loop counts per task, and stop the run when a limit trips. KPMG found 43% of organizations have implemented usage or token budgets. In a multi-agent system, a runaway loop multiplies, because each agent can trigger calls in the next one down the chain.
Test the system, not just the agents. Evaluate full workflows, including failure cases: what happens when one agent returns nothing, returns nonsense, or times out. Our AI orchestration work starts from these failure paths, because that is where multi-agent systems differ most from a single, well-instructed agent.
Decide who owns it after launch. A multi-agent system that nobody on your team understands is a liability. If an outside partner builds it, settle documentation, knowledge transfer, and ownership up front, and plan for the people who will maintain it. We covered that in who owns a vendor-built AI agent.
Common Questions About Multi-Agent AI Controls
What is a multi-agent AI system?
A multi-agent AI system is a setup where several AI agents, each with its own role, tools, or data access, pass work to one another to complete a task. One agent might gather information, another analyze it, and a third act on it. The design can help with complex work, but it adds handoffs, cost, and points of failure.
How many companies use multi-agent AI?
KPMG's Q3 2026 AI Pulse found 25% of large US organizations actively developing or implementing multi-agent systems, up from 6% in each of the prior two quarters. The survey covered 314 US leaders at organizations with $1 billion or more in revenue, between July 24 and August 25, 2026. Mid-market adoption was not measured.
Are multi-agent systems better than a single AI agent?
Not by default. A June 2026 study, "The Illusion of Multi-Agent Advantage," found that automatically designed multi-agent systems consistently underperformed a simple single-model baseline while costing up to 10 times more. Multiple agents make sense when the work needs genuinely different tools, data, or permissions, not as a default architecture.
What controls does a multi-agent AI system need?
At minimum: a separate identity and least-privilege access for each agent, structured and logged handoffs with end-to-end tracing, human approval for irreversible actions, token and loop limits with automatic shutoff, and testing of full workflows including failure cases. In our experience, building these in at design time costs far less than adding them after an incident.
If you are moving from one agent to several, or inheriting a multi-agent system you cannot fully trace, talk to the Golabs team. We will start by asking whether you need more agents at all, and then help you build the controls that let the ones you keep run safely.