This is a response to a Linkedin Post
I think the field for building Agentic AI systems lends itself to technologists with a wide array of skills. These systems use programmatic glue to chain and orchestrate multiple AI calls and systems together. Tools I've needed so far is a fine understanding of Python (Jupyter notebooks is nice) and access to an LLM. The specific AI skill I've leveraged most is inspecting and evaluating the output and updating my system accordingly. I've read quite a few research papers to help learn some of the design patterns being researched as well. I haven't deep dove the different models and have focused mostly on chaining and orchestration challenges.
Day to day:
Make an "Agent" (LLM call with user message and system message)
Save the output to a file, inspect that file (see if I like it or not). If no iterate on the prompts.
Make another Agent. Pipe the output from the last step into the new agent. Inspect new file output.
Find bug, refactor previous 2 agents, make third because I was asking too much from the last 2.
Once I understand what's goodish. Make evaluating/critiquing agent. If evaluations are good you can automate this pipeline by asking the evaluator if the task is complete.
Get a couple good pipelines and then start working on orchestrating between the pipes.
Oh, while you're doing all this it becomes more difficult to understand the system by looking at the text file outputs so you're going need to constantly tune your logging (observability) system so that you can determine Signal vs Noise.
Top comments (0)