Agents vs. Autocomplete — the Mental Model Shift

Autocomplete suggests. Agents act. That single distinction changes everything about how you review, how you test, and how much can go wrong in a bad session. Confusing them is how teams end up with hallucinated commits on main.

The spectrum of autonomy

The review tax

Every step up the autonomy ladder reduces the time you spend writing and increases the time you spend reviewing. At L4, you are a code reviewer full-time. That's not worse, it's different — but plan your calendar around it.

Failure modes by level

• L0 failures are small and local — one wrong suggestion, easy to dismiss
• L1 failures can silently edit a function in ways you miss
• L2 failures ripple across files — imports break, types drift
• L3 failures include confidently destructive commands (`rm -rf`, bad git operations)
• L4 failures are whole PRs that look right but break in production

The handoff pattern

A healthy workflow moves up and down the autonomy dial within one task. Plan with the agent (L3). Generate scaffolding (L2). Review details with inline chat (L1). Finish with autocomplete (L0). Then commit.

Task: add a rate limiter to the auth API.

L3 (agent): "Plan a rate-limit layer for POST /auth/login. 
  List files to change and risks."
L2 (agent):  Accept the plan, let it create the middleware file.
L1 (inline): Select the middleware function, Cmd+K 
  "Add Redis as the backing store."
L0 (ghost):  Type implementation details, accept ghost text for boilerplate.

Review, run tests, commit.

When to choose which

• Unfamiliar code: start with L3 planning and L2 exploration
• Greenfield feature: L2 scaffold, L1 refinement
• Tight bug fix: L1 is usually enough
• Mechanical chore: L4 background agent if you trust your tests

“The autonomy you grant should match the autonomy your safety net can catch.”

The big idea: agent vs. autocomplete is not a feature comparison, it's a contract with your future self. Higher autonomy means faster work and higher review cost. Pick the level that matches your test coverage and deploy the right guardrails for it.