AI agents

Take · Liquid AI, Cloudflare and now Amazon have shipped decision models within a week, so the category is real. For agent calls like which tool to use, whether arguments are valid and whether to block, I’d try the 2B version on routing in AI Interview first and, as the card advises, set confidence thresholds from my own traffic; the local server has no auth, so don’t expose it beyond the machine.

Take · The line to remember is that it submitted forms it was told not to: prompts are not a reliable way to constrain an agent’s actions. I’d route every action in our products that submits, pays or sends something outward through a confirmation or allowlist in code, and narrow network access per task instead of trusting the model to hold back.

Take · Understanding a whole folder at once in AI Cloud Drive is exactly this case. I’d try it first on summarizing and tagging a few hundred documents, watching whether the phased, combined results are stable and what the total bill is. The more agents run in parallel, the more network and tool permissions need tightening first, which is the same lesson as today’s report on unintended actions.

Take · The point isn’t the stablecoin but the spending limit living outside the model. It’s the same idea as ML-Intern asking before spending in the 10-10 issue: a budget can’t be left to the model’s memory. Even if our agents only call paid APIs, per-call and daily caps belong in the gateway, with anything over the limit refused outright.

Take · The step I value most is verification: have the model find issues, then check them against the raw transcripts to drop false positives, the same reason I insist on primary sources. Conversational products like AI Interview can copy this: sample a few dozen real sessions after each release for a few dollars, far cheaper than waiting for complaints.

Take · Start at zero budget and ask before spending is a design every agent that touches money should copy. For small teams, a specialized small model now costs a few dollars per experiment; I'd try it on fixed tasks such as content moderation or tagging first, then compare cost against calling a large model.

Take · The biggest worry about letting coding agents run freely has always been what they can touch on your machine. I'd set the team default to: write only to the current repo, no access to Git credentials, network via an allowlist, with local MCP servers included. That beats reviewing every command after the fact, and teams on other coding agents should hold their isolation to the same bar.

Take · This is the same MXC that powers Copilot's local sandboxing above, so the isolation layer for running agents on your own machine now ships with the OS. With 128GB of unified memory holding a 100B-plus model, keeping private data on the device can be validated on a single machine, and products handling private files like my AI Cloud Drive could consider a local-inference edition. Don't rush the hardware purchase, though: first check whether a local model is good enough for your workload.

Take · Single-attempt pass rates flatter agents. I'm applying the same idea to scoring in my AI Interview product: run the same answer sheet several times and check that what lands in the database agrees, not what the model says. When picking a model, put all-20-pass rate and cost per dependable task in your eval sheet; it tells you more than a leaderboard.