If the scaffolding does not matter, what does?

More agents do not make an AI system reliable. A new study says so, and it leaves one question open: then what does?

You have seen the slide. Twenty agents in boxes, a manager agent on top, a price per box. You are asked to buy it before anything has booked a single entry.

What the slide costs you

Every box is a licence and an integration. When the result is wrong, twenty boxes mean nobody can say which box was wrong, so nobody is accountable. And the agents still cannot reach your accounting system, so they guess, confidently, from a copied spreadsheet.

What the study found

In September 2026, researchers at EPFL and Apple tested the slide. ¹ They built the plainest agent they could: one session, a strong model, direct access to read, write and run things, and a time budget. It matched or beat four published multi-agent systems on the same tasks: a medal on 62.5 percent of the competitions against 47.1 percent for the best of them. Then they added what the slide recommends. Sub-agents did not help. Three agents talking to each other made it worse, from 55.7 percent to 33.3 percent.

Their conclusion: the leverage is in the model and the runtime it is given, not in the scaffolding around them.

We accept it. The limits are in the footnote. The direction is clear enough to act on.

Then what does matter?

One agent, with five things. In this order.

1. Access. The agent reaches your systems and files itself, through access you grant and can withdraw. The study supports this one.
2. A source on every answer. Every figure traces to its record. When there is no source, the agent says so. This is a data foundation, not a model feature.
3. Guardrails it cannot pass. “Never pay an invoice” in a prompt is a rule, and a rule can be argued with. Access that cannot make a payment is a guardrail. Access rights, read-only connections and code enforce the limits.
4. Checks on real work. Not benchmarks. Does the posting match the invoice? Does every figure show its source? How many corrections this month, and did each one change the rules?
5. A person who knows where they sign off. One agent, one accountable operator: the person at your company who checks and signs off the knowledge workers’ work. Per job: acts alone, asks first, or the operator signs off. A step moves to “acts alone” only after it has proved itself, and it can always move back. Approving every line by hand is retyping.

The study’s one clear result, that agents coordinating with each other did worse, is the case for number five: the agent collaborates with a person, not with other agents.

One step this week

Take the proposal on your desk. Count the agents. Then count the systems they can reach, and find the sentence that says where a person signs off. If the first number is the only impressive one, send it back.

If you want a second pair of eyes on it, that is what a conversation with us is for.

Footnotes

1. Brilliantov, Hernández-Cano and Abbé, “How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?”, arXiv 2609.40303, 30 September 2026, arxiv.org/abs/2609.40303. Medal rates are self-selected submissions at a 24-hour budget on machine-learning competitions: Table 15 with GLM 5.2 (Malena 62.5 percent, AiScientist 47.1 percent), Table 1 for the additions (base 55.7 percent, sub-agents 45.2 percent within the confidence interval, three parallel agents with a broadcast tool 33.3 percent, outside it). On two smaller models other systems led by mean within overlapping intervals, so the authors claim “no worse than”. Malena cost about six times more to run than one of the systems. Limits stated by the authors: machine-learning benchmarks only, possible benchmark contamination, possible tuning asymmetry, and inter-agent coordination left to future work. Read on 24 October 2026. ↩