An InfoQ article traces how agentic LLM systems moved from linear prompt chains to graph-based workflows and then to harnesses that execute code. It recommends validating model outputs and picking the simplest design that meets reliability, speed and cost needs.
Launched in August by SpaceXAI, the Grok Bot agent books, buys and subscribes on the user's behalf, handing back control only to pay. JDN tested it on three everyday tasks.
Stockholm-based Strawberry has taken its AI browser out of beta with version 1.0. Rather than living in a chat window, its assistant works inside the browser with access to the tabs, tools and accounts people already use.
Cloudflare has released Auto Router in public beta via AI Gateway; a classifier scores each request for complexity, ambiguity, stakes and context dependence to choose a model. Cloudflare's own benchmark puts it ten points below Claude Opus 5.5 on success rate at a third of the cost, though practitioners question how stakes are judged over a long agent session.
A JDN column warns that if IT departments are the ones paying for AI agents, companies have a return-on-investment problem they have not yet noticed.
When customers' AI agents start negotiating with companies' agents, the question will not be whether they can, the column argues, but how much they are allowed to decide.