Enterprise teams building multi-agent AI systems can pay a computational premium for gains that do not stand up under equal budget conditions. A new Stanford University study finds that single-agent systems match or outperform multi-agent architectures in complex reasoning tasks when both are given the same reasoning token budget.
However, multi-agent systems come with additional computational overhead. Because they typically use longer reasoning traces and multiple interactions, it is often unclear whether their reported gains are due to architectural advantages or simply greater resource consumption.
Stanford University researchers to isolate the real driver of performance compared single-agent systems with multi-agent architectures on complex multi-hop grounding tasks under equal "thought sign" budgets.
Their experiments show that, in most cases, single-agent systems match or outperform multi-agent systems when computation is equal. Multi-agent systems gain a competitive advantage when a single agent’s context is too long or corrupt.
In practice, this means that a single-agent model with an adequate reasoning budget can provide more efficient, reliable, and cost-effective multi-hop reasoning. Engineering teams should reserve multi-agent systems for scenarios where single agents reach a performance ceiling.
Understanding single and multi-agent partitioning
Multi-agent frameworks, such as scheduling agents, role-playing systems, or debate groups, decompose the problem by having several models operate in partial contexts. These components communicate with each other by transmitting their responses.
Although multi-agent solutions show strong empirical performance, comparing them to single-agent baselines is often an imprecise measure. Comparisons are heavily muddled by differences in test timing. Multi-agent setups require multiple agent interactions and create longer reasoning traces, meaning they consume significantly more tokens.
ddAs a result, when a multi-agent system reports higher accuracy, it is difficult to determine whether the gains are due to better architectural design or the cost of additional computation.
Recent studies show that when the computational budget is fixed, the developed multi-agent strategies often underperform compared to robust single-agent baselines. However, these are mostly very broad comparisons that do not take into account nuances such as the difference between different multi-agent architectures or the difference between operational and constitutive traits.
“The main point of our paper is that many comparisons between single-agent systems (SAS) and multi-agent systems (MAS) are not apples-to-apples,” paper authors Dat Tran and Douwe Kiela told VentureBeat. “MAS often achieves more efficient test timing through additional calls, longer traces, or more coordination steps.”
Revisiting the multi-agent problem under tight budgets
To create a fair comparison, the Stanford researchers set a strict “thought mark” budget. This metric monitors the total number of tokens used for intermediate reasoning only, excluding the initial request and final output.
The study evaluated single- and multi-agent systems on multi-hop reasoning tasks, that is, questions that require combining many different pieces of information to arrive at an answer.
In their experiments, the researchers observed that single-agent setups sometimes stop their internal reasoning prematurely, leaving the available computing budget unspent. To avoid this, they introduced a technique called SAS-L (longer-thinking single-agent system).
Instead of switching to multi-agent orchestration when the model fails early, the researchers suggest a simple quick and budget change.
"The engineering idea is simple," Tran and Kiela said. "First, restructure the single-agent query so that the model explicitly spends its available reasoning budget on pre-response analysis."
By instructing the model to detect ambiguities, list candidate interpretations, and test alternatives before providing a final answer, developers can recover the benefits of collaboration within a single-agent setup.
Their experimental results confirm that single agent is the most powerful standard architecture for multi-hop reasoning tasks. It produces the most accurate responses when consuming less compelling cues. When combined with Google’s dedicated models like Gemini 2.5, the longer-thinking option produces better overall performance.
Researchers rely on a concept called “Information Processing Disparity” to explain why a single agent outperforms a swarm. Multi-agent frameworks present unique communication bottlenecks. Every time data is aggregated and transferred between different agents, there is a risk of data loss.
In contrast, a single agent thinking in one continuous context avoids this fragmentation. It maintains access to the richest representation of the task and thus becomes more data-efficient within a fixed budget.
The authors also note that businesses often overlook the secondary costs of multi-agent systems.
"What enterprises often fail to appreciate is that orchestration is not free," they said. "Each additional agent introduces communication overhead, more intermediate text, more opportunities for lossy generalization, and more room for error aggregation."
On the other hand, they discovered that multi-agent orchestration is superior when an agent’s environment is mixed. If an enterprise application must handle highly degraded contexts such as noisy data, long inputs full of distractions, or corrupted data, the single agent struggles. In these scenarios, structured filtering, decomposition, and inspection of a multi-agent system can recover relevant information more reliably.
The study also warns of hidden evaluation pitfalls that falsify multi-agent performance. Relying solely on the token counts reported by the API severely misrepresents how much computation the architecture actually consumes. Researchers found these accounting artifacts while testing models like Gemini 2.5, proving that it is an active issue for enterprise applications today.
"For API models, the situation is more complicated, as budgeting can be opaque," the authors said. They advise developers to reliably evaluate architectures "record everything, measure visible justification traces when possible, use provider-reported justification-token counts when exposed, and treat those numbers with caution."
What it means for developers
If a single-agent system matches the performance of multiple agents under equal reasoning budgets, it wins in total cost of ownership by offering fewer model calls, lower latency, and simpler debugging. Tran and Kiela warn that without this foundation, "some enterprises may pay a large ‘too much tax’ for architectures whose apparent advantage comes from spending more computing than actually thinking more efficiently."
Another way to look at a decision boundary is not how complex the overall task is, but where exactly the bottleneck is.
"If it’s mainly depth of thinking, SAS is often enough. If this context is fragmentation or degradation, MAS is more defensible," Tran said.
Engineering teams should stick with a single agent when a task can be managed in a single consistent context window. Multi-agent systems are necessary when an application handles highly degraded contexts.
Looking forward, multi-agent frameworks will not disappear, but their role will evolve as boundary models improve their internal reasoning capabilities.
"The main takeaway from our paper is that the multi-agent structure should not be considered as a default assumption that more agents automatically mean better intelligence, but as a purposeful engineering choice for specific bottlenecks." Tran said.
