Part of Mechanisms of Vibe Coding

Mechanism: Pair Vibe

One-sentence definition

Pair vibe is two-role collaboration: the human steers direction, goals, and checkpoints while the agent executes edits and commands — like pair programming with the model as the fast typist, not the author of intent.

The problem

Solo vibe: you disappear into agent output and accept the path because it is moving. Or you micromanage every line and lose speed. Pair vibe without roles becomes either abdication or fight — agent drives product choices, or you type manually while AI idles.

Explicit roles fix both.

Symptoms:

• You cannot explain why the approach was chosen
• Agent picks architecture while you only watch
• You rewrite agent code by hand without updating prompts
• High-risk work with no human navigator

How it works

  Human (navigator):
    · goal, constraints, scope fence
    · approve plan / mode switches
    · checkpoint, calibration tier
    · stuck detection → escalation

  Agent (driver):
    · edits, commands, loops within approved bounds
    · propose options, not pick product direction
    · surface verify output

Session rhythm:

1. Human sets goal block.
2. Agent proposes plan or executes slice.
3. Human runs or watches verify, reviews diff.
4. Human says next direction — not open-ended “continue.”

Escalation to full human driver on high calibration tier.

When to use it

• High-risk or unfamiliar domains
• Learning a codebase while shipping
• Teaching junior devs agent workflows
• When solo sessions show vibe drift or completion bias

When not to use it

• Low-risk repetitive transforms with strong automation — human navigator optional
• Navigator disengaged “while agent works” — that is abdication, not pair vibe
• Two humans without agent role clarity — different pattern (human pair)

Failure modes

Abdication — “Fix it” without goal; agent navigates. Fix: Human owns goal and checkpoint.

Micromanagement — Human edits every line; agent as autocomplete only. Fix: Bigger PR-sized slices with trust after verify.

Role swap mid-session — Agent reframes product; human accepts silently. Fix: Mode switch; human approves plan.

Pair without verify — Driver says done; navigator never ran tests. Fix: Navigator runs verify.

Minimal example

Context: Verbose health slice 2 — bootstrap uptime.

Pair session:

• Human: Goal + constraints + “propose bootstrap trace only, no edits yet” (debug mode).
• Agent: Trace notes + hypothesis.
• Human: “Edit bootstrap.ts line X only; run test.”
• Agent: Edit + show diff.
• Human: Verify locally, checkpoint, merge or reject hunks.

Done when: Human can demo the fix without reading agent chat — they steered, agent executed.

Tool instances (optional deep-dive)

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

Cursor

• Human on keyboard for verify and git; agent in Composer for edits.
• Voice or shared screen: navigator does not Accept All without diff review.
• Plan mode = navigator time; Agent = driver time after “approved.”

Related mechanisms

• Escalation — human takes wheel at Level 5
• Mode switching — roles per mode
• Confidence calibration — high tier → pair vibe default

Try it yourself

Exercise: One session as explicit pair vibe — you may not let agent start edits until you approve a one-paragraph plan. Agent may not declare done until you run verify.

Done when: You wrote two sentences: what you steered vs what agent executed.