Dev Log: AI Agents
I've been experimenting quite a bit with the AI Agent stack. While many people have been suggesting things like CrewAI and LangChain, I decided to go down to the root level to work on an AI Agent that does some work contextualizing and then editing files in a file server.
The Stack
Initial Challenges Finding a framework that works with MCP is difficult, as it setting up and connecting to the server initially. It feels less like plugging in Legos, as it does a really solid Arduino board. It definitely isn't for complete beginners, but with a bit of work you can do it.
There's also many agent frameworks that have been put out that claim to be easy to setup, but are poorly maintained or no documentation and then have been thrown out after a few months
On Building Agents Yourself Initially, I started building Agents myself with no framework, just a basic client per the MCP documentation. This approach is fine, but the ability to stand up agents quick using a framework cannot be understated. There's also a huge downside to just using a client where you can have your Claude talk to a server or do it from a chat box. Most of the power and magic of AI comes from feeling like there is an autonomous agent.
Final Thoughts
AI agents are here, but they're not fully autonomous. More thoughts as I get further, but there is a substantial amount of chaining and prompting that almost feels like crafting a workflow in Zapier than setting up an AI and letting them loose. I'm sure this will improve, but right now the name of the game is how do you have the human managing the AI without making it feel like they're managing an employee who really doesn't want to be there and is putting in half effort on the whole thing.