I read this: http://ampcode.com/how-to-build-an-agent.

I followed along and pasted the code and it is nice to see how, underlying all the hype around “agents”, the essential concept is very simple (and very powerful):

While a “conversation” is the underlying primitive of LLMs, one can easily imagine building different UX based on the specific use case.

I then took some time to delve more deeply into A2A (TLDR: it’s a protocol for agent-to-agent communication, still early and I don’t think I have seen much actual adoption yet) and MCP (a standard for “exposing tools and resources to LLMs”).

Specifically, I watched this https://www.youtube.com/watch?v=m3IE6JygT1o. Interesting notes:

Some other MCP stuff I’ve seen and intend to spend more time to explore more deeply:

I am convinced that we will see many many MCP servers shortly.

Meanwhile, I have started on a new side project to build a quick AI agent using Cloudflare’s Agent SDK. Look forward to reporting back once that’s ready.


Came across this on evals by Andrew Ng

https://x.com/andrewyng/status/1912908679344693711?s=46