Today’s radar was heavy on money: a wallet for an agent, token prices down 80%, and a frontier model fitting on a single GPU. I started the day thinking about spending limits and ended it thinking about unit economics.

In brief

Cloudflare announced programmable stablecoin wallets that let AI agents buy APIs, data and content directly over HTTP. A person keeps a main wallet and creates separate virtual wallets for each agent, with a spending limit, approved suppliers and a cap per transaction. The product is not available yet: funding and spending authorization still have no date. But the design points to an important conversation for credit: if an agent can spend within a pre-approved limit, it can also operate subject to a credit limit—with governance defined before, rather than after, a loss.

The Cloudflare announcement stood out because it has the strongest implications for anyone working on the product side of credit.

Anyone who works on credit products knows that the hard part is almost never the decision. It is the payment: who pays, how much, under which limit, who authorizes it, and what happens when something goes wrong.

That is why this news caught my attention. Cloudflare announced programmable stablecoin wallets that let AI agents buy APIs, data and content directly over HTTP. A person keeps a main wallet and creates separate virtual wallets for each agent, with a spending limit, a list of approved suppliers and a cap per transaction.

Notice the design. The agent is not asking for approval at every purchase. The human defines the envelope beforehand, and the agent operates inside it.

This is precisely the vocabulary of credit: limits, authority levels, approved counterparties and exception policies. The difference is that the party making the spending decision is now software running on its own at three in the morning.

It is also the type of decision that calls for the fundamentals of AI agents: a clear scope, defined tools and an autonomy limit proportional to risk. The AI governance guide helps turn these controls into operations; the AI risk matrix provides a starting point to classify impact; and the AI agent evaluation template organizes quality, cost, latency and oversight before rollout.

This is where I see the most interesting opportunity. If an agent can spend within a pre-approved limit, it can also be subject to a credit limit. Consumption-based micropayments, automatic reconciliation of a large volume of small transactions, machine-generated receivables. These are problems the receivables industry already knows how to solve, but at a scale and granularity that did not exist before.

It is worth keeping our feet on the ground: Cloudflare has only opened name reservations for now; funding and spending authorization still have no date. This is a directional announcement, not a finished product. But direction matters, because it shapes what will make it into roadmaps over the next two years.

I am optimistic about this kind of move. It forces the governance conversation to happen at the start, in the design of the limit, rather than after the loss.

For anyone curious to read the full announcement, here is the link: Cloudflare wallets for agents.

The rest of the radar

OpenAI cuts GPT-5.6 Luna pricing by 80% — an order-of-magnitude drop in inference cost reopens previously unviable use cases and puts pressure on companies monetizing AI by seat or credit. Read more

Mistral’s Shieldstral: 3B open-weights multimodal moderation — a safety layer configurable through natural-language policy, without retraining and cheap to run, which unlocks AI features in sensitive domains. Read more

DeepSeek V4 Flash running on a single AMD MI300X — a frontier model with 1M context fitting on one accelerator changes the unit-cost equation and makes self-hosting viable. Read more

Maple-Preview: a 20B ternary MoE at 120 tok/s on an iPhone — fast on-device inference removes per-token cost, network latency and much of the privacy risk. Read more

The Warp Agent CLI — coding agents stop being tied to an IDE and become portable infrastructure, accelerating prototyping and technical discovery. Read more

LLMs reward expertise — AI productivity gains are proportional to the operator’s expertise, affecting ROI expectations, onboarding and the UX design of assisted features. Read more

Cloudflare applies engineering standards with AI — a real case of AI in an internal process with a clear quality metric, which can be replicated to justify automation investment. Read more

When AI Benchmarks Plateau — if public benchmarks have saturated, selecting a model by leaderboard becomes a poor decision and evaluation needs to be tailored to the use case. Read more

AI adoption in product management: the 2026 data — a market benchmark on how peers are using AI and which success metrics have already changed. Read more

That is it for today. More tomorrow.