Amelia Wattenberger, designer at Augment Code and former GitHub Next researcher, argues that AI agents are inverting the difficulty curve of software development. Starting a project is now trivially easy. Finishing one is becoming the hardest problem in the field. Her thesis: prototyping just got cheaper, but the last 30% of any project, the part requiring taste, judgment, and coherent direction, got significantly harder.
Wattenberger is designing Intent, Augment Code's agent product, which treats a workspace as its core primitive instead of a chat thread. That architectural decision is the crux of the episode. The conversation covers the tradeoffs between one-worktree-per-agent and one-worktree-per-task, the full arc from autocomplete to chat to CLI and back to purpose-built UI, and why developers are experiencing a genuine identity crisis as agents take over keyboard time. GitHub Copilot is referenced roughly 12 times as the inflection point that started this era.
Read the full transcript or listen to understand why the workspace-versus-thread distinction matters for how agents actually maintain context across long-running tasks. The infrastructure detail is also worth your time: the discussion of Incus, ZFS snapshots via Proxmox, and sandboxed agent environments explains the concrete engineering bets Augment is making. The identity question Wattenberger raises, what developers are actually for now, does not get a clean answer. That is the point.
[READ ORIGINAL →]