This matches a lot of what I have seen while building my tool for AI. The speed boost is real, but the confidence gap is the part people underestimate. AI can get you to looks correct very fast, but that’s not the same as is correct — especially once the task touches real logic, context, or architecture. Treating it like a junior developer is probably the healthiest mental model I’ve found too: useful, fast, often impressive, but still something you need to review, guide, and verify.

Our view is that it has to be both. Event-based tracing is great for reconstructing the path: tool calls, inputs, outputs, transitions. Snapshots are better for inspecting state at meaningful boundaries and seeing what actually changed. Once agent workflows get long enough, events alone usually aren’t enough for debugging. You need state checkpoints too. So the mental model we keep coming back to is events for sequence, snapshots for state and both for real traceability.

Agree. This is very close to what I’ve seen while building Origin. Once you connect AI to tools, files, and workspace state, it becomes much more of a system design problem than a model problem. Usually the first failures I notice are bad context and broken state handling, not the model itself. That’s also why traceability matters so much, if you can’t see what changed and what caused it, debugging turns into guesswork.