(2026-05-18) In Late2024 The First Version Of The Mcp Spec Was Launched

Christian/curious: In late 2024, the first version of the MCP spec was launched and my friend @jlehman_ showed me what was probably one of the top three most mind-blowing demos of AI I've ever seen — he was using Claude Desktop and, if I remember correctly, a local MCP server called something like "what could go wrong" to control his whole computer just by prompting Claude.

It's now taken about a year and a half for those core components to mature, but here we are.

I have a similar intuition right now about primitives like the /goal command, the /loop command (for outcomes, as it's known in the Anthropic Managed Agents API), and critically the forthcoming Dreaming feature — combined with turnkey deployment platforms like Claude Managed Agents that afford you the ability to not only scale up a number of coin-operated computers and point them at a particular salient problem, but most interestingly, finally start to get compounding returns through automated mining of the agent traces generated through massively parallelized work at a particular task.

This is verifiable in practice. Andrej Karpathy's auto-research experimentation is probably the generalized version: spin up ten or a hundred or even a thousand agent threads to tackle a particular issue

The Dreaming technique and synthesis of organizational memory is going to blow people's minds

And finally — this is where I think it's going to force economic norms to fundamentally change, or at least raise some pretty radical questions — when you can do this across team members or an entire organization. A recurring job that distills everything down and makes the common knowledge pool (into which everyone is implicitly contributing by making their traces legible to the org administrator) more useful. That automated creation of memory through synthesis of the work people are doing across a team is genuinely novel territory, and I think will trigger more fat-tail events in 2027.