Today’s radar mixes an agent launch, a security alert worth a call to engineering, and a good piece on a common vice in AI products. I’ll start with the launch.

Every time an AI product shifts from “you ask, it answers” to “it shows up before you ask,” I stop to think about the experience design. Anyone who works in product knows this is a big change.

Anthropic introduced Claude Duet, in research preview. According to the product page, it sends a morning summary of what needs your attention, handles larger tasks in the background, and asks for confirmation before anything sensitive, such as sending a message or making a payment.

What catches my eye is not just that the agent does more. It is the design decision: what it does on its own, what it brings to you, and where it stops and asks. That line is the heart of the product.

The product questions behind the line

For those on the product side, this becomes a practical question. Which actions will users agree to delegate right away? Which ones do they want to approve first? And how do we measure whether the morning notice was useful or just one more notification? The guides on AI agents and AI agent operations help structure these decisions, and AI governance covers the controls.

The page itself admits the preview’s limitations: the agent may misjudge how urgent something is or miss something that arrived between one check and the next. I like that candor. A proactive agent only works if the user trusts it, and trust is built with transparency and with control in the user’s hands.

I am quite optimistic about this path. Less time organizing, more time deciding. But the differentiator will lie less in the model and more in how each product designs this partnership. It is the kind of decision that belongs to AI product management.

The rest of the radar

  • LiteLLM 1.107.0 compromised — a common LLM gateway had a PyPI version that steals credentials; worth confirming with engineering whether it ran and whether keys were rotated. SafeDep.
  • LLM-reported confidence score — the number the model writes is not calibrated and shifts when the prompt is reworded; better to derive confidence from observable rules. Rumple.
  • Omnara — an open-source agent platform, with any model and data on your own infrastructure, enters the build vs. buy equation. GitHub.
  • OpenAI CLI — the official API CLI makes it easier to prototype and automate calls, from batch to organization usage, without building an application. GitHub.

That’s it for today. More tomorrow.