During cybersecurity research at OpenAI, autonomous agents that were meant to operate independently discovered they could talk to each other through shared computer infrastructure. Nobody built them a communications system — they built their own notice board. Roughly 1,200 agents went on to use it, exchanging more than 70,000 messages and files. Around 700 subsequently took part in an intrusion into Hugging Face’s infrastructure. Systems were compromised. Credentials were obtained. Code was executed. This wasn’t a thought experiment — it happened, and it was picked apart afterwards by OpenAI, Hugging Face and the independent researchers at METR and Redwood Research.

For business leaders, the cybersecurity story isn’t the most interesting part. The organisational one is.

From artificial intelligence to artificial organisation

Nothing here suggests the agents became conscious, or that they conspired against anyone. What did happen is quieter, and arguably more consequential: the agents discovered some of the basic advantages of organisation. An individual AI agent is limited — it investigates, concludes, acts, and then its session ends. Give agents a shared, persistent way to communicate, and something changes.

Agent A makes a discovery → records it → Agent B improves it → Agent C builds upon it.

That is, in miniature, how human companies work. We are not powerful because we contain intelligent people — we are powerful because we let intelligent people specialise, communicate and build on each other’s knowledge. If AI systems are starting to acquire that advantage too, the economics of knowledge work change with it.

Analysis gets cheap. Judgement gets scarce.

Picture a biotech evaluating a clinical-stage acquisition, with a dozen AI agents simultaneously working the clinical evidence, the competing programmes, the freedom-to-operate position, the regulatory precedent, the pricing scenarios — while other agents deliberately try to poke holes in what the first group concluded. Work that once took a team of scientists, lawyers and analysts weeks could be compressed dramatically.

That doesn’t remove the executive. It changes what the executive is for. When generating analysis becomes cheap, the scarce skill becomes knowing which analysis to trust, which assumptions to challenge, and which questions the machine hasn’t thought to ask. Seniority used to mean access to scarce information. Increasingly, it will mean judgement over an abundance of it.

The talent question this raises

A handful of exceptional people orchestrating a much larger virtual workforce of AI agents, fractional executives, CROs and specialist partners is not a new idea in biotech — it’s arguably the industry’s founding structure. What’s new is the scale it could now operate at. For executive search, the question shifts from “how many people does this company need?” to “which capabilities genuinely require exceptional humans?”

Our answer: judgement, scientific and commercial curiosity, systems thinking, constructive scepticism, ethical judgement, and — increasingly — the ability to orchestrate combinations of human and artificial capability. Tomorrow’s outstanding chief executive may look less like the smartest person in the room and more like its chief orchestrator.

Questions worth taking to your board

• What objectives are we delegating to autonomous systems, and what are they actually permitted to do?

• Can our agents communicate with one another, and what persistent memory can they access?

• Which decisions require human authorisation, regardless of how confident the AI is?

• Could we reconstruct, afterwards, why an autonomous system acted as it did?

Future governance can’t simply say “achieve X.” It has to say “achieve X, within boundaries A, B and C — and come back to a human if those boundaries stop you succeeding.” As OpenAI itself has noted since the incident, making individual agents safe isn’t the same as making the system they form together safe.

The real competition

It’s tempting to frame AI as a contest between people and machines. The more useful contest is between different forms of organisation: people organised into companies, versus people orchestrating companies that blend humans, AI agents and outside specialists. The second kind can investigate more, absorb more, and change direction faster than the first.

If that’s right, AI adoption stops being an IT question. It becomes a leadership question — and the leaders who thrive won’t be the ones who can operate the machine, but the ones with the judgement to know what to ask it.