Friday, 19 June 2026

The infrastructure for AI agents is quietly becoming real. Over the past week, we've seen Google, Microsoft, and GitHub move from talking about autonomous agents to actually standardizing how they discover and use tools online. That shift from concept to specification matters more than it might sound.

The Agentic Resource Discovery spec is the piece I find most interesting here. It's not flashy—it's a technical standard for how agents find, verify, and communicate with available tools. But standards are how technologies move from labs into production. You need common ground before hundreds of companies can build agents that work together reliably. By the way, this is exactly the kind of unglamorous infrastructure work that usually precedes a technology wave. Nobody gets excited about HTTP until suddenly the web is everywhere.

Microsoft is moving fast on the product side. Copilot Cowork is designed to handle complex business tasks by orchestrating multiple tools—Excel, Outlook, and other Microsoft 365 applications. This isn't a chatbot that gives you advice; it's supposed to actually *do* work. Charles Lamanna's team is positioning this as delegation, not assistance. The distinction matters. And I think they're onto something real here, assuming the execution holds up.

There's a pricing signal worth noting too. Microsoft just shifted Copilot from flat-rate to usage-based pricing because power users were driving compute costs too high. That's telling. It means the system is being used intensively enough that the old model broke. It also suggests that agent-based systems, which orchestrate multiple API calls per task, will have different economics than simple chat. Usage-based pricing might become standard for agentic systems.

Now, the harder question: are these agents actually reliable? Anthropic's latest paper on readable AI activations is relevant here. They've made progress on interpretability—understanding what's happening inside the model—but readability and faithfulness aren't the same thing. You can understand what a model is doing and still not trust it to handle a critical business process unsupervised. That gap between readable and reliable is where the real work still lies.

On the creative side, xAI's Grok Imagine Video 1.5 launching at 86 percent below Sora's pricing is a reminder that the AI video market is still consolidating. Price competition this aggressive suggests margins are under pressure and customer acquisition is the game. But that's a different conversation from the agent infrastructure story, which feels like it's entering a new phase of maturity.