@rida This is the right direction. Worktrees + better runtime ergonomics reduce friction, but the remaining gap is knowledge shape.
Agents still break when docs are blob-like instead of traversable structures.
Most agent failures I see are not model failures.
They're interface failures:
raw blobs, weak structure, no stable paths.
Give agents a filesystem they can reason over and reliability jumps.
Hot take: "parse PDF" is table stakes now.
The real bottleneck is giving AI agents documents in a format they can reliably operate on.
I'm focused on agent-ready filesystems + bash-native workflows.
I'll share what works (and what breaks) in production.