Good luck slowing this down
Ben's Bites
VCs are hyped on Instinct, but Ben says the personal agent feels slow and a bit bossy. He’s also wary of Meta knowing too much and of AI agents acting before you do.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ben’s latest rundown has one clear thread: the personal-agent wave is arriving, but the experience still feels rough around the edges. He tried Instinct, the agent VCs have been talking up, and came away underwhelmed. The problem wasn’t just that it tried to be proactive. It was the vibe of the thing. It felt like extra work to keep it on track, and at one point it even sent a message saying it had finished meeting follow-ups and that he should reply if he wanted the details. That’s not exactly a warm welcome.
There’s a bigger concern underneath the annoyance. These systems want more context, more memory and more access to your life, but they don’t always make that visible by default. Ben says he’d rather be able to see and edit what an agent remembers, instead of trusting a black box to quietly build a dossier. He’s also not thrilled about the idea of handing more personal data to Meta, which he calls a shaky privacy partner.
That scepticism sits alongside a broader prediction: every AI company is going to ship its own personal agent, just as every product now seems to come with an AI chat box glued to it. The market may be moving in that direction no matter what users think. But if the first impression is “this feels like a boss,” that’s a problem, not a feature.
The rest of the newsletter reads like a snapshot of the next phase of the AI race. Dario Amodei has written about slowing the frontier for safety. Sam Altman says OpenAI won’t do an IPO in 2026 because of the work ahead on alignment, control and safety. Meanwhile OpenAI is pushing out more product pieces, from GPT-Live-1 in the API to an Agents API, while Claude Code adds a way to test whether plugins actually help. The pace is not slowing. It’s just getting more organized.
Ben also flags a bunch of smaller signs that agent tooling is spreading fast: Pi now lets users bring their own API access differently, Bolt Forge is packing in more models, Replit has Routines for recurring work, and there’s a new ChatGPT mini widget for desktop. The theme is obvious. Everyone wants agents to feel useful. Very few of them, right now, seem to know how to make that feel calm.
My take — AI-written commentary, not fact-checked reporting
The industry keeps shipping agents like autonomy is the product, when most people just want a tool that stays out of the way. Making software that talks first and explains later is a great way to create tiny office managers nobody asked for. Visibility and control are the moat here, not the wiggle and chatter.
Read more about this at: Ben's Bites