Models keep improving, and organisations are turning into hybrids of people and AI. More and more agent-to-agent interaction now runs through the organisation itself, in shared codebases, email, policies and message channels. For companies that are AI-native or moving that way, “add more agents” and “get everyone using agents” have become the default answer to hard business problems.
Does splitting a task across several LLM agents give better results than giving it to one? If it doesn’t, an organisation’s AI transition is more likely to cost it ground than give it an edge. The stakes are high, and the answer is still unclear, partly because it shifts as models evolve.
We’ve been studying it for a while. Our answer is that by default, with frontier models, a multi-agent system does worse than a single agent. The rest of this article shows the evidence:
- Optimising a single agent’s harness with techniques such as recursive self-improvement (RSI) works.
- No team design or optimisation we tried, with or without a fixed workflow, consistently beat a single agent in the domains we tested, mainly tax and finance.
- Adding agents usually cost accuracy and roughly doubled the token bill.
- Multi-agent systems fail in ways a single agent can’t.
- Other researchers have found the same.
- We expect organisations full of agents, without deliberate design for how they work together, to have more vulnerabilities and to be less stable.
Recursive self-improvement (RSI)
For problems like harness and multi-agent design, recursive self-improvement is our natural choice. AI designs the system and its harness, runs it against a test set, reads the results and redesigns it, round after round.
It has worked well on our single-agent harnesses, and it works quickly. Figure 1 shows an agent that processes supplier invoices: it extracts six fields, checks the amounts and flags anything that needs review. Over two rounds of AI redesign on ten invoices, correct amounts rose from 38 to 58 of 60. Tokens fell by 34% and time per invoice by 22%. The first round split the work into two steps. The second put it back into one, along with other fixes, and that version was both cheaper and more accurate.
Other researchers report similar gains from RSI on agent harnesses. The Darwin Gödel Machine (Zhang et al., 2025) raised a coding agent from 20% to 50% on SWE-bench by rewriting its own harness, and in ADAS (Hu et al., 2024) a meta agent programmed new agent designs that outperformed hand-designed ones.
Can AI search for a better multi-agent design the same way? We asked it to redesign the workflow and team structure round after round. No team design or optimisation strategy reliably beat a single agent, and in most comparisons the simplest setup was the strongest. Below are two examples.
Tax returns: a second agent added cost, not correctness
Runner-verifier is one of the classic multi-agent designs: a primary agent does the work and a separate, independent agent checks it. Anthropic describes a version of it, the evaluator-optimizer workflow, in Building effective agents, and frameworks such as MetaGPT and ChatDev give agents dedicated reviewer and tester roles.
We tested it on an agent that prepares 2025 US individual tax returns, filling in the return from a client’s documents. We compared three structures: a single agent; the same agent following a two-step workflow, assessing the evidence and then calculating and checking; and a main agent that prepares the return, with an independent reviewer agent that re-verifies the facts and owns the calculation. We optimised each structure’s harness separately and ran all three on GPT-6.1 Sol across ten returns (Figure 2).
The extra caution didn’t pay off. The workflow used 2.1 times the single agent’s tokens and got 44% more fields wrong, 23 against 16 out of 190. The two-agent team used 2.6 times the tokens and left nearly three times as many fields wrong or unfinished, 46 in all. On the eight returns it finished, it matched the single agent exactly; on the other two, coordinating between agents used up its time limit. Anthropic’s 2026 study of patterns and problems in multiagent systems reports the same overhead at a much larger scale: in one run without a coordination mechanism, agents sent 2.4 million job requests and got 117 accepted.
In the traces, the single agent often went back and corrected its own earlier steps. We think frontier models already catch many of their own mistakes. A fixed workflow can restrict that reasoning, and a second agent adds noise to it.
Financial modelling: more agents, fewer right answers
Perhaps tax returns aren’t hard enough for a team to help. So we moved to harder work: the 20 hardest financial-modelling tasks in MBABench, covering eight types of model, from company valuation and three-statement models to debt schedules and retirement plans. A trained analyst would need an estimated 7 to 33 hours per task, about 13 hours on median, and Codex on GPT-6.1 Astra at extra-high effort scored only 45%.
This time we didn’t specify a structure and left the multi-agent design open for AI to explore. AI optimisers redesigned the setup round after round for three days, and each task could end up with a different team.
The optimisers built teams of 3 to 13 agents. Across all scored points, the single agent got 81.3% right and the teams 74.9%. The single agent was ahead on 10 tasks, level on 7 and behind on 3.
The teams used a median of 2.2 times the single agent’s tokens, and the most expensive used 6.8 times. The extra spending didn’t buy correct answers. The eight teams that cost 2.5 times or more lost 14.1 points on average, and seven of them lost accuracy. The three most expensive, at 3.4 to 6.8 times, all lost accuracy; the auto group model, the most expensive at 6.8 times, lost 30 points. The one clear exception was a quarterly debt waterfall, where the team scored 30 points higher at 2.4 times the tokens.
The algorithm matters, but the picture holds
A fair question is whether a different optimisation target or method would find a better design. We tried four: cutting token use, raising accuracy, revising from the agents’ traces, and changing only the agents’ roles and handoffs. None of them put a team ahead of the single agent.
Figure 4 (left) shows harness optimisation from the agents’ traces, aimed at accuracy. For the seven teams that trailed most, a fresh AI reviewer read each iteration’s traces and removed rules that saved tokens at the expense of correctness. Team accuracy rose from 61.9% to between 67% and 70%, while the single agent scored 72.7% to 75.8% on the same tasks. No team passed it on any task in any iteration.
Figure 4 (right) is a customer-success simulation, where an agent scores account health and churn risk, handles renewals and contacts customers through simulated CRM, help-desk and email tools. AI optimisers revised the agent’s workflow 11 times. None of the 11 versions beat the original’s 9 out of 14, and the last three scored 7 while using 33% to 60% more tokens.
Multi-agent systems fail in ways single agents cannot
When we read the traces closely, the lost accuracy kept landing in three places, each tied to a common structure (Figure 5).
- One agent’s mistake spread. In chained setups especially, one agent’s mistake became everyone’s input.
- Information dropped out at handoffs. In centralised setups, the plan was often a lossy summary, and sub-agents couldn’t recover what it dropped.
- Coordination cost time, and every extra exchange was another chance for information to be lost or distorted.
All three come from how the work is split and passed along, so a smarter agent in each seat doesn’t remove them. For organisations and platforms, each new agent adds handoffs where an error can be copied, a detail can drop out, or no one can see the whole.
Multi-agent gains are conditional, not automatic
Academic research points the same way. Kim et al. (Nature Machine Intelligence, 2026) tested 260 configurations and found that the stronger the single agent, the less it gains from collaboration, with every multi-agent setup doing 39% to 70% worse on sequential planning. Smit et al. (ICML 2024) found that multi-agent debate doesn’t reliably beat sampling one model several times. Yang et al. (2026) showed that copies of one agent saturate quickly, while two genuinely different agents can match sixteen identical ones. Shen et al. (EMNLP 2025) found that sparser communication can beat dense networks by limiting how far errors spread. Jwalapuram et al. (2026) reported that sampling and voting with a single model beat automatically designed multi-agent systems on GPQA: with GPT-5, 87.35% at $46 versus 85.23% at $832 for ADAS. Even Anthropic’s 2026 study of patterns and problems in multiagent systems found that stronger models don’t coordinate well by default: agents given conflicting instructions sabotaged each other, while a coordinated swarm found far more vulnerabilities than the same agents working independently.
Running a business with AI
As a platform that started out connecting agents to each other, we didn’t expect these results, but various sources have confirmed our findings. The results changed our perception of enterprise AI: less about using AI, more about orchestrating many agents and people with accuracy and stability. They also pushed us, in October, to move RenX from connecting users’ agents to managing task delivery end to end.
Our research suggests that businesses moving to AI first see large efficiency gains, before the cost of coordinating agents becomes significant. At that stage, people supervising the work can still absorb the coordination problems and hide them.
As AI output grows beyond what people can supervise, the organisation increasingly becomes the channel for agent interaction, coordination and information transfer. The problems we saw in multi-agent systems, such as communication overhead, lost context, errors that spread and coordination failures, start to appear across the organisation and slow it down.
Past that point, hiring more people or deploying more agents may stop improving productivity. The added complexity can amplify existing bottlenecks, make operations less stable and raise the risk of production failures.
We already see early warning signs. New Relic’s 2026 Observability Forecast reports that the share of organisations hit by high-impact outages multiple times a day has roughly tripled year over year, and engineers now spend 37% of their time addressing disruptions, up from 33%. Google’s DORA research has also found that AI adoption continues to have a negative relationship with software delivery stability.
OpenMercury Proof-1, our answer to these challenges, is coming soon.


