Part of Mechanisms of Vibe Coding

Mechanism: Over-Automation Trap

One-sentence definition

The over-automation trap is automating agent loops or hooks without strong goals, verify, and termination conditions — so wrong assumptions get repeated faster and at scale.

The problem

You wire an agent to “fix CI until green” on every push. It patches tests to match bugs, adds suppressions, or refactors unrelated code — fifteen times overnight. Automation removed the human pause that would have caught nonsense.

More automation without better control patterns is amplified vibe drift and completion bias.

Symptoms:

• CI green but product broken
• Test suite weakened by auto-fix loops
• Quota burned with no mergeable outcome
• Fear of enabling agents after one runaway night

How it works (trap vs healthy automation)

  Trap:
    vague goal + auto loop + weak verify
         │
         ▼
    faster wrong commits

  Healthy:
    explicit goal + verify hook + termination cap + human checkpoint on merge

Safe automation minimum:

1. Named verify command (not “lint stuff”)
2. Max iterations + timeout
3. No auto-merge to main
4. Constraints in loop prompt template
5. Alert human on cap hit

When to watch for it

• Before unattended agent loops
• “Run until done” scripts without exit criteria
• Hooks that invoke full Agent on every failure
• Replacing stuck detection with more iterations

When automation is appropriate

• Verify-only hooks (tests, lint) without agent
• Agent on failure with fixed prompt file and caps
• Playground branches only

Failure modes

Automate before spec — Loop without examples or AC. Fix: Spec first, automate second.

No termination cap — Infinite loop. Fix: Hard N and wall clock limit.

Auto-merge — Agent PR merges itself. Fix: Human checkpoint always.

Verify theater — Agent edits tests to pass. Fix: Human review + mutation testing awareness.

Production automation — Live fix loops. Fix: Staging/playground only.

Minimal example

Context: Nightly “fix health.test.ts” agent job.

Trap version: Runs until green, any diff, auto-commit.

Safer version: Max 3 iterations; fixed goal/constraints prompt; commit only to auto/health-fix branch; notify human; no merge without review.

Done when: Runaway job cannot touch main or weaken tests without you noticing.

Tool instances (optional deep-dive)

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

Cursor

• Loops with iteration limits; no auto-accept all.
• Hooks: verify scripts default; agent hook optional and sandboxed.
• Review .agent/ or branch output each morning before trust.

Related mechanisms

• Termination condition — required for automation
• Verify loop — must be real, not rhetorical
• Hooks / automation — design carefully

Try it yourself

Exercise: Audit one automated or loop workflow. List what could go wrong if it runs 50 times with a wrong assumption.

Done when: You added a cap, verify-only step, or human gate.