Part of Mechanisms of Vibe Coding

Mechanism: Branch-per-Experiment

One-sentence definition

Branch-per-experiment means each agent attempt, slice, or hypothesis gets its own git branch — so you can compare, abandon, or merge one outcome without entangling others.

The problem

One branch accumulates three approaches: Redis cache attempt, in-memory fix, and the actual solution. Commits are interleaved. You cannot merge the winner without cherry-picking through noise. Deleting bad work means surgery.

One branch for all vibes is one shared scratchpad — convenient until it isn’t.

Symptoms:

• Branch name does not match current task
• git log reads like a novel, not a PR
• Hard to show colleague “approach A vs B”
• Abandoning work feels expensive

How it works

  Hypothesis or slice
       │
       ▼
  branch experiment/
       │
       ├── merge (winner)
       ├── abandon (delete branch)
       └── compare (diff against base)

Naming: experiment/health-uptime-bootstrap, agent/verbose-health-slice-2 — searchable, disposable.

Pair with rollback: branch tip is the snapshot.

When to use it

• Best-of-N comparisons
• Escalation retries that should not pile on same diff
• Split & conquer slices that might merge independently
• Any “let me try something else” moment

When not to use it

• Single linear slice already on a feature branch — one branch per feature is enough
• Branch explosion without cleanup — delete abandoned experiments weekly

Failure modes

Branch hoarding — 40 stale experiment/*. Fix: Delete or archive; document winner.

Same branch, new experiment — Commit approach B on approach A’s branch. Fix: New branch from base.

Merge without compare — First finished branch wins. Fix: Verify + checkpoint against criteria.

Base drift — Experiments branch from old main. Fix: Rebase or branch from current target.

Minimal example

Context: Two ideas for uptime: read from process.uptime() vs custom startedAt.

Steps:

1. experiment/health-uptime-process — agent implements A, test.
2. experiment/health-uptime-started-at — fresh branch, agent implements B, test.
3. Compare curls and diff size; merge one; delete the other.

Done when: Loser branch deleted without touching winner’s history.

Tool instances (optional deep-dive)

Portable idea above; this section is tool-specific. Date: June 2026.

Cursor

• Agent on current branch; you create branch before agent starts.
• git worktree for parallel local experiments without switching.

Other tools

GitHub compare view, git diff main...experiment/foo.

Related mechanisms

• Playground — branches as playground units
• Rollback — reset branch to snapshot
• Best-of-N — N branches for N attempts

Try it yourself

Exercise: For your next “try another way” moment, create a new branch from base instead of continuing on the current one.

Done when: You can delete the failed branch with git branch -D guilt-free.