Small agency, large team? What a team of AI assistants does and does not deliver

• A team of AI assistantsAI Assistant An AI assistant is a chat program such as ChatGPT, Claude, Gemini, or Perplexity that you can ask a question in natural language and receive a written answer.More about this term works better on tasks you can break down into parts, and worse on work that must be done step by step or where a single agentAI Agent An AI program that independently performs a series of tasks, such as checking pages or conducting measurements.More about this term already performs well.
• Research identifies most errors in setup, collaboration and oversight, not in the language modelLanguage Model The computational engine behind an AI assistant that has learned how language works.More about this term itself.
• In my view, who checks the work matters more than the size of the team.
• If I am unavailable, two backupsBackup A backup copy of your website, allowing you to restore it if something goes wrong.More about this term are in place, and another agency can take over the websites, though not the swarm.
• Want to know where you stand? Take the free scan: https://eqwise.nl/gratis-website-scan/.

An objection I hear: better a large agency than a sole trader. Behind that objection lie two questions: who covers if that single person is unavailable, and who checks their work? A team of AI assistants appears to offer an answer. In the industry, such a team is called a swarm. In a swarm, a coordinator distributes the tasks, specialists handle their respective parts, and separate reviewers verify the results. However, ‘team’ sounds grander than it is, and the question is what research actually shows. I review five studies. They examine AI systems and people working with AIAI Computer programs that perform tasks normally requiring human intelligence, such as writing text or answering questions.More about this term, not agencies.

On work that you can break down into parts. Anthropic, the creators of the ClaudeClaude The AI assistant from Anthropic.More about this term models, reported on 13 June 2025 that a team with a lead agent and sub-agents excels particularly at queries with multiple independent directions, such as finding the board members of every IT company in the S&P 500, where a single agent failed with slow, sequential searches. In an internal test by Anthropic itself, this setup scored 90.2% better than a single agent. The creator was therefore measuring its own product. The setup used roughly 15 times as many tokens (the unit in which a language model processes text) as a standard chat conversation. On another benchmarkBenchmark A comparative figure, such as the average within your industry.More about this term, BrowseComp, which measures how effectively agents retrieve hard-to-find information, token consumption accounted for 80% of the variance in scores. Anthropic itself concludes that such a team works primarily because it applies sufficient compute.

Google Research tested 180 agent configurations in December 2025. For financial analysis that can be divided into parts, a team with central coordination scored 81% better than a single agent. That is a different test and a different comparison from Anthropic’s, so the figures cannot simply be combined.

For people working with AI, the gain also depends on the task. Dell’Acqua and colleagues (Harvard Business School and BCG, Organization Science, March 2026) gave consultants 18 realistic tasks within the boundary of what AI can handle. Using AI, they completed 12.2% more tasks, 25.1% faster and with higher quality, compared to consultants working without AI. That study concerns a single human using AI, not a team of agents.

A team of agents performs worse on work that must be done step by step. The Google study across 180 configurations showed that on sequential tasks, such as planning, every multi-agent configuration scored 39 to 70% worse than a single agent. If a single agent already scores above roughly 45%, additional agents yield diminishing or even negative returns. Anthropic also highlights a limitation: according to Anthropic, most programming tasks contain fewer genuinely parallel subtasks than research work. A team is therefore not an enhancement you can apply everywhere.

A similar boundary exists for people working with AI. Dell’Acqua and colleagues found that when consultants tackled a task outside what AI can manage, they were 19% less likely to produce a correct solution than without AI.

By measuring. METR, a non-profit organisation that evaluates frontier AI models and agents, ran a randomised controlled trial in July 2025 with experienced open-source software developers. Using AI tools from early 2025, they took 19% longer. That concerns time, not the likelihood of a correct solution. The developers themselves believed they had been faster. The actual slowdown likely lies between 2% and 39%. This relates to one specific type of task at a single point in time. What feels faster is not always faster: have the actual returns of an approach measured instead of relying on intuition.

Research shows that errors stem mainly from how a team is organised. Cemri and colleagues (UC Berkeley, NeurIPS 2025) investigated seven open-source multi-agent systems. Those systems failed in 41% to 86.7% of cases. Most failures occurred in how the system was configured, in collaboration between agents, and in oversight, rather than in the language model itself. The Google study of 180 setups found that configurations without central oversight propagate errors more than those with central oversight.

My reading: whether at an agency, large or small, or within a team of agents, who checks the work matters more than the size of the team. None of the five studies in this article is about agencies. For you as a client, the primary question is therefore: does someone who did not produce the work review it?

I direct and run the swarm myself. The swarm is the methodology behind SEO and GEOGEO GEO stands for generative engine optimisation: increasing the likelihood that AI assistants such as ChatGPT recognise and mention your business when someone asks them for advice.More about this term, the work that ensures you are found on Google and mentioned by AI platforms. It consists of over a hundred agents, twenty of which are reviewers. The swarm enables me to work more effectively. In my experience, that is because I have embedded thirty years of knowledge into the agents and because they review one another. That is my practical experience, not an empirical measurement. Strategy I handle myself, based on thirty years of experience. That aligns with the Google study of 180 configurations: for sequential tasks, every team scored worse than a single agent. I continue to learn and draw support from my agents and from two people outside the swarm: a colleague and an external website developer.

Standard review is handled within the swarm and by me. The twenty reviewers, who are agents themselves, verify the work of the other agents, and I always evaluate the final output myself. Occasionally, I ask experienced colleagues for their view on work produced by the swarm, though not routinely. An example from this article: an early draft attributed Anthropic’s finding that token consumption explains 80% of score variance to the wrong test. It belonged to BrowseComp, not to the test with the 90.2% result. Source checking caught the error, and it was corrected before the article went live.

Two people provide cover for me, and another agency can take over the websites I build, though not the swarm itself. The primary backup is Elza van Swieten, a colleague with thirty years of experience who covers during my holidays. The second is my external website developer, with whom I have worked for over twenty years. Together, they can, in my assessment, resolve any issues with the websites. I build websites to standard specifications so that another agency can step in. The swarm and the knowledge embedded within it are run solely by me; no one else can take them over, including those two individuals.

The outcome at OmniHealth resulted from the entire EQwise project and does not serve as standalone evidence for the swarm. That project involved work on their websites and discoverability. From January to September 2026 inclusive, compared with the same period in 2025, revenue increased by 51% and bookings by 45%, both measured within OmniHealth’s booking system and adjusted for the measurement period. This is an outcome of the EQwise project, delivered entirely with AI support and partly through the AI swarm. Most of the growth occurred prior to 11 September 2026, when I permanently assigned the swarm to the SEO and GEO work.

Want to know whether your site is prepared for search in Google and mentionsMention A place where your business name is mentioned, with or without a link.More about this term by AI platforms? Request the free scan at https://eqwise.nl/gratis-website-scan/. The scan is free and carries no obligation, and you will receive a report within two working days. It evaluates seven areas: technical foundation, SEO, GEO, speed, trust, UX and conversionConversion The moment a visitor does what you hope for, such as calling, filling out a form, or making a purchase.More about this term. You will see where you stand in Google, how AI assistants describe your brand, whom they mention, and which three improvements will deliver the greatest return. I will then show you what can be divided and what cannot.

1. Anthropic, ‘How we built our multi-agent research system’, 13 June 2025: https://www.anthropic.com/engineering/multi-agent-research-system
2. Google Research, ‘Towards a Science of Scaling Agent Systems’, December 2025: https://research.google/blog/towards-a-science-of-scaling-agent-systems-when-and-why-agent-systems-work/ and arXiv 2512.08296: https://arxiv.org/abs/2512.08296
3. Cemri et al. (UC Berkeley), ‘Why Do Multi-Agent LLM Systems Fail?’, NeurIPS 2025: https://arxiv.org/abs/2503.13657
4. METR, randomised trial with experienced developers, July 2025: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
5. Dell’Acqua et al. (Harvard Business School and BCG), ‘Navigating the Jagged Technological Frontier’, Organization Science, March 2026: https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838