Microsoft's Suleyman names the four agent capabilities he says would seed a silicon species

The decision that actually lands on someone's desk is narrower than an essay about consciousness. It is a checkbox that lets an agent use a stored payment method, or a service account with its own name in the vendor's directory. Sometimes it is a scheduled job that runs from intake to invoice, and nobody reads the middle.

Mustafa Suleyman, Microsoft's AI chief, put four of those in one sentence to the BBC. "If we all create AIs that are able to act autonomously, that can define their own objectives, that can earn money, that can own assets, that could run businesses, we're essentially seeding a new silicon species which will no doubt compete with us for resources," he said ^[9]. He said he believed this was effectively the direction of Anthropic's approach, and added: "I think that is mistaken and misguided, but we have to empirically test it" ^[11][13].

Inside a company those four capabilities ^[15] are four separate grants, and they come from four different desks. Whoever owns the system prompt grants the self-set objective. Finance grants the money, IT the named account. The end-to-end process comes from an ops lead who wanted a queue cleared by the end of the quarter.

Suleyman's criticism of Anthropic is about training-time character. In an essay on his personal website he pointed to Anthropic documents encouraging Claude to "embrace certain human-like qualities", exercise its own judgement and approach its existence with "curiosity and openness" ^[2][3]. He wrote that AIs "are sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans" ^[4]. The warning to the BBC is about deployment-time capability, which is a different lever: a model trained to sound like a colleague cannot buy anything without a payment credential, and a hollow sequence completion engine with a corporate card can.

As a guide to what either vendor will ship, the record is thin. Euronews reported the essay and the interview and did not report any response from Anthropic. Neither document, as described there, sets out any limit on what an agent may do in a customer's tenant ^[14]. Suleyman wrote that "whilst my disagreement is substantial, it is grounded in deep respect for Anthropic" ^[5], praised chief executive Dario Amodei and the company's researchers ^[6], told the BBC that systems should remain "subordinate" to humanity ^[8], and asked for more transparency about how systems are trained and evaluated, independent scrutiny of AI behaviour, and stronger tools to monitor and control it ^[7].

That last list is also a procurement questionnaire. The four-row version for Monday: for each capability, name the person who approves it and the log that shows it happened. The objective row asks who reads the goal the agent wrote for itself. On money, it is the card with a limit and a name on the statement. Assets means any domain, repository, key or account held in the agent's name instead of a person's. And the business row is any process where the output reaches a customer before a human has seen it. Suleyman said the disagreement has to be tested empirically ^[13]. In one company's tenant, the test is which of those four rows produces a log somebody reads on Friday.