I have been reading about AI Agent software architectures and designs for work and I have to confess I'm fascinated by them. The idea of a loop that takes an action, potentially makes a tool call, performs (or doesn't!) some check to determine if a goal was reached and returns to the user is extremely valid. It's a way of using a computer that I think speaks to a lot of people. The only problem is, well, all of the LLM and execution of the loop! 😆
1 Boost
0 Likes
The most tantalizing idea though is that you could define an agent as any program that runs this REAct loop. Meaning you could have an agent that doesn't revolve around an LLM but something else. Keep the user paradigm swapping out the planet killing techno-fascist underpinnings and you end up with something that could be a really useful computer tool.
MCP is fascinating. What if we exposed workflows programmatically instead of building blocks to construct those workflows ourselves? In some ways it reminds me of GraphQL where you build the whole query at once instead of piecemeal.
At Dayjob we have a different standard protocol we deveolped that's been used in industry for about 10 years now. Turns out that MCP has extremely similar underlying pattern. It's been odd watching the AI industry appropriate our terms and then see it all spread.