The Card on File
Sheel Mohnot tried to pay a medical bill with two personal agents. One refused. "Muse wouldn't do it because it only has my tokenized credit card data. Instinct stored my credit card data in plain text, which is not great" The less careful agent got the bill paid.
Mohnot, of Better Tomorrow Ventures, uses Instinct all the time. He is also tired of the setup. "I actually have personal agent fatigue, um, where I'm pretty reluctant to try another one and again, input all of my data." His fellow investors are stricter. Chris Farmer of SignalFire: "there's only so many points of entry into my private data that I want" Kanu Gulati of Khosla Ventures keeps hers on errands. "I haven't used them for any of cases where I would have to share any personal information like my credit card or SSN."
OpenAI's Tibo Sottiaux states the trade from the seller's side. "the utility that you get from these systems is kind of capped by, you know, how much access it has" His advice is to ration it anyway: "you start with limited access." The agent gets more useful exactly as fast as you get more exposed.
Inside companies, the answer is a log with a name on it. Kevin Mandia, who founded Mandiant and now runs Armadin, runs attack agents for a living. "Like we log every single thing our agent does. Source IP address, time of date and what it did." His standard for everyone else: "we all clearly have to mature into it in a way where we have the accountability when these things go rogue."
Manik Surtani, co-founder of the Agentic AI Foundation, finds the sign-off already hollow. "Even though it's being signed off by an author, that person's not even reading that code, so why then should someone else go and review that code?" He would let AI do the review, on one condition, and then asks the question nobody has answered. "And how do you hold that person accountable to that?"
Jeff McMillan, who ran AI at Morgan Stanley, puts the name on the builder. "That is the person who builds the tool owns the deliverable and owns the outcome." Humans cannot sit in every loop, so "you have to have human above the loop".
Andy Doyle, Kantar's chief people and agent officer, has found the weak point in that arrangement. Staff question an agent. Call it something else and the scrutiny drops. "Yet when we see digital workers in our environment. People are a bit more, oh, the digital worker told me it was true."
Last week the question was whose name goes on the agent. This week it is whose card.
Sources: Interviews from This Week in Startups (Oct 7, Sheel Mohnot of Better Tomorrow Ventures, Chris Farmer of SignalFire and Kanu Gulati of Khosla Ventures), Masters of Scale (Oct 10, Tibo Sottiaux of OpenAI), a16z (Oct 6, Kevin Mandia of Armadin), Dev Interrupted (Oct 6, Manik Surtani of the Agentic AI Foundation), Metrics that Measure Up (Oct 6, Jeff McMillan of McMillan AI), and The Modern People Leader (Oct 9, Andy Doyle of Kantar).
Practitioners say personal and enterprise AI agents get more useful as they get more access, which puts a named person on the hook for what they do.