Agent Learning Roadmap

Why this page exists

Plenty of people have already taught this better than I could, and most of it's free. This is just the specific path I'm taking through it, starting with tool-use reliability specifically, since it's the most concrete failure mode to go find in the wild, ahead of things like agent evals or multi-agent coordination that only make sense once that's solid. Tracked here so I actually follow it instead of leaving it in twenty open tabs. If you're following along too, this is the actual order I'd suggest, not just a byproduct of me tracking my own progress.

The journey

Three phases, worked in order. Phase 1 is the wedge itself, the thing I'm actually practicing. Phase 2 is the tool I'm using to practice it, leveled up as I go. Phase 3 is the knowledge a PM background doesn't hand you for free, picked up alongside the other two rather than front-loaded. None of the three need to be "finished" before the next one starts. They run in parallel, just in this order of priority when something has to give.

Phase 1: The wedge, in practice

I picked tool-use reliability as my wedge because it's the most concrete failure mode there is. An agent either calls the right tool with the right arguments, or it doesn't. No vibes, no "it seems smart," just a pass or a fail I can actually go find. These four stages are how I'm working through it for real, not just reading about it: build something small enough to break, learn exactly how it breaks, write it down before I forget why it mattered, then go check my thinking against people who've shipped this stuff for real. Each stage leans on the one before it.

0 / 0 done

Phase 2: Claude Code feature adoption

The tool I'm using to execute Phase 1. Every feature, ranked by how soon it should land in your workflow.

0 / 0 adopted

Phase 3: LLM production knowledge

The knowledge gap alongside Phases 1 and 2, not before them. A self-study curriculum, tiered by how foundational each topic is.

0 / 0 done