Instinct raises $1 billion at a $10 billion valuation for an agent that operates users' phones and computers

Instinct's agent is built to use the phones and computers people already own, so users do not learn a new interface ^[4]. Its website lists the connections: email, messaging, the user's screen, audio and location ^[4]. The company describes the goal as a personal agent that performs tasks autonomously ^[18].

Every connection on that list belongs to the user. An agent that works through a screen reaches a booking site the same way its owner does ^[4]. Pressure on app makers to publish scoped, machine-facing interfaces would follow if Instinct moved to direct integrations with them. Its described design goes through the device instead ^[4]. For now the boundary that counts is the grant a user gives Instinct, and RuntimeWire's account sets the bar for it: a booking, message or purchase made on a user's behalf must be correct, authorized and recoverable when it is not ^[15].

Shinn's research is the clearest clue to how he means to meet that bar. As a Northeastern student he co-developed Reflexion, a framework in which a language model checks its own output and uses the feedback to improve later attempts ^[10]. Northeastern reported that the method reached 91% on the HumanEval coding benchmark, 11 percentage points above the result then attributed to GPT-4 ^[11]. I think separating the doing from the checking is the right starting point for an agent that books tables and places calls ^[5]. The 91% is still a claim about someone else's workload. Code can be run again, and a failing run tells the checker what broke. A placed phone call is harder to take back than a failed test. For the number to transfer, a wrong booking would need an equally clear failure signal and an undo.

Checking also costs calls. Each later attempt is another trip to the model, and RuntimeWire notes that an agent may already need repeated calls to models and external services to finish a single request ^[16]. The Information's user report also described demand straining capacity ^[6], and RuntimeWire has since covered slowdowns and capacity warnings ^[9].

The capital is there to absorb that load for a while. The Series C is four times the size of the $250 million Series B that TechCrunch reported in August ^[2]. Added to the $350 million TechCrunch counted as total funding at that point, it brings reported funding to about $1.35 billion ^[3]. The announcement does not include user or revenue figures ^[7].

A user alleged that the assistant surfaced another person's financial details ^[17]. The screenshots cited in that report did not establish whether another user's information had been exposed or the assistant had generated the details ^[17]. The first would be a permission failure, and the second is the kind of output a self-check loop exists to catch. Shinn denied a data breach ^[17].