7 sessions on bubbletea with a coding agent: the agent failed at least once in 3, and 2 finished without a human stepping in.
Takeaway: I'd start by pointing the agent straight at the specific subsystems I care about (permissions, memory/compaction, sub-agents) instead of letting it wander the whole repo tree first — that would have gotten to the comparison faster.
Takeaway: Next time I'd point straight at the debug HTML dump and cookie state instead of starting from the UI layout code — the real bug was three layers deep in auth cookie filtering and stale selectors, not rendering.
Takeaway: Next time I'd point straight at internal/browse/view.go and mention the shadowed layout variable as a likely suspect, since that's what actually broke the counter.
Takeaway: I'd start the first prompt by naming the exact command shape I want (`economist demo`, no flags) and flagging up front that I want a clean DataSource abstraction so future section/rendering changes don't create debt — that's what I ended up asking for mid-session anyway.
Takeaway: Next time I'd tell the agent upfront to verify file existence and rebuild/reinstall the binary before claiming a fix works, since most of this session was chasing 'it's not implemented' and stale-binary ghosts rather than the actual feature.
Takeaway: Next time I'd start by asking for a design/research doc first and only prototype code once that's reviewed, since the first TUI attempt got built too fast and had to be thrown away.