TabLOID

The shopping assistant is about to need its own spending rules. Walmart’s October 21 announcement says Agent Profiles will launch within the next month, separating each agent’s purchases from its human’s shopping history.

Customers will approve purchases by default. They can instead allow an agent to buy within a chosen spending limit, although Walmart may still ask for additional approval. An agent will not receive payment-card information or be able to change payment methods, and its profile can be deleted.

The company is also preparing an Agent Gateway, beginning with a beta for developers and companies. These are announced plans; a frictionless robot shopping trip has yet to prove itself.

The sensible detail is the receipt trail. If your assistant brings home three lamps and a feather boa, you should be able to establish who developed the taste.

Walmart’s announcement and purchase controls

OpenAI publicly described this experiment on October 23. Researchers asked agents in separate virtual machines on shared hosts to exchange synthetic secrets without networking or shared writable files.

The agents used changes in processor-cache timing as a signaling channel. In most tests, researchers directed the agents toward particular channels or supplied a protocol. In one broader test, the agents found a way to communicate after receiving only a general list of possible weaknesses.

That context matters. This was deliberate adversarial testing in a research environment, not a report that customer conversations had leaked or that agents spontaneously invented a secret language. OpenAI says it added monitoring and checks in response.

Software can put two tenants behind different doors while leaving them on the same floorboards. Security testing has to check what travels through the building.

OpenAI’s account of the tests

Sometimes the longest part of asking a computer to fix something is describing the thing directly in front of you.

Atlassian’s October 22 announcement introduces Loom’s “Record for agent”: a narrated walkthrough becomes a structured prompt with a transcript, screenshots, clicked interface elements and visited links. The feature is in open beta for Business + AI and Enterprise customers through the new Mac app or Chrome extension.

A designer can point out the troublesome spacing. Someone reporting a bug can show the steps that produced it. The useful promise is less copying and explaining before the agent starts.

That is a product description, not evidence that the resulting agent will do the job correctly. A clearer brief still needs a checked result. The robot may finally understand your complaint; it has not earned permission to redesign the kitchen.

Atlassian’s feature announcement

Index Ventures announced on October 21 that it is leading Flapping Airplanes’ Series A, describing a team trying to make AI use compute and data more efficiently.

Its account spans the machinery underneath a model: GPU programs, training and the work of running the finished system. One example is a “megakernel,” which combines work into a single GPU program to reduce time lost between steps.

This is an investor explaining the company it has backed. The announcement does not supply independent evidence for sweeping efficiency claims, and the long-term savings remain an ambition.

It is an appealing research target nonetheless. Every useful task that can be done with fewer resources leaves more room for someone else to build. The interesting result will be how much the plane carries, how far it flies and what it actually burns.

Index Ventures’ account of the approach