I appreciate Chris’ “glass half full” attitude about the current state of LLM dev tools and Elixir’s prospective place amongst them.

Also love the hack of revising the agent’s “memory” to omit failures that could otherwise poison the context. I have definitely had to abandon chat threads when the assistant became fixated on implementing the wrong solution because it appeared in the conversation more often than the correct solution.

Strongly agree that the current frontier of agentic workflows is who can figure out the best way to convince the chatbot to stay on task, and that the folks developing the higher-profile products (Claude Code, etc.) are not that far ahead of the rest of us in that area