Concurrent agent workspaces

Stub. Full step-by-step playbooks will land here. For the why, read Git concurrency enshittification.

Goal

Run human work and one or more agents on related changes without corrupting a single shared index, without surprising path breaks, and without multiplying dependency trees needlessly.

Near-term options (choose deliberately)

Approach Prefer when Watch out for

Native worktrees

You need real separate checkouts and standard Git CLI everywhere

Absolute path drift; duplicated node_modules / build dirs; ref lock contention

jj / op-log style tools

You want working-copy-as-revision and less index pain

Team Git interoperability and tooling familiarity

GitButler virtual branches

You want one path and hunk lanes for unfinished work

Native Git commands outside the tool desync workspace state

Process namespaces + CoW clones (DIY Phase A)

You control the agent runner and can pin a stable /workspace

Operational complexity; still not FS-native editions

Hard rules for agents

  1. Do not let agents write into the human interactive working tree by default.

  2. Prefer ephemeral sandboxes that export patches / revisions rather than mutating HEAD in place.

  3. Share content-addressable dependency stores (e.g. pnpm) across sandboxes when possible.

Phase B pointer

Filesystem editions (connectome-fs) make “new workspace” an O(1) edition fork with stable paths. Until then, this how-to stays in userspace tradeoffs.

Related