Smell #4: Over-Delegation

Severity: Critical

Over-Delegation: Granting excessive autonomy to AI assistants or agents, effectively abdicating engineering judgment and oversight to a machine.

Symptoms

How to recognize this smell in your process:

  • [ ] You use "Background Autonomy" modes for complex architectural changes.
  • [ ] You let agents work "overnight" and merge their changes in the morning without individual review.
  • [ ] You prompted something like "Improve the whole project" or "Refactor all components."
  • [ ] You feel like the AI is "running the project" and you are just the "PR approver."

Self-Assessment

If you checked even 1 item, you are in the Critical Risk Zone for Over-Delegation.

Example

The "Morning After" Nightmare

Scenario: A developer gives an agent background access to "fix all styling and performance issues." Morning Result: 127 commits.

  • The "Fix": The agent refactored 15 files, deleted 3 "unused" files (that were actually critical), and updated 5 dependencies.
  • The Debt: Two production-breaking silent failures and three critical security holes.
  • The Cost: It takes the team 3 days to revert and reconstruct the logic.

Debt Impact

Over-Delegation is the only smell that contributes to all seven categories of synthetic debt. It is catastrophic.

| Debt Category | Impact | |---------------|--------| | 🏗️ ARCH | Total loss of structural coherence. | | 🧠 KNOW | No human understands the 100+ automated changes. | | 🔐 SEC | Agents often remove security checks to "fix" bugs. | | ⚖️ IP | Zero IP protection for unsupervised output. |

How to Fix

  1. Immediate Stop: Disable all background autonomy modes immediately.
  2. Revert & Review: If you have merged unsupervised AI code, revert to the last known "Human-Understood" state.
  3. Forensic Audit: Conduct "Digital Archaeology" on any automated changes you choose to keep.

How to Prevent

  • Mandatory Checkpoints: Agents must stop and wait for human approval after every modification.
  • Atomic Tasks: One agent session = one specific, small engineering task.
  • The "Senior Reviewer" Mindset: Treat an agent like a high-speed junior intern who needs 100% supervision.

Related Smells

Book Reference

  • Chapter 5: Agents Unleashed — how autonomy begins to multiply debt.
  • Chapter 8: Background Autonomy — the full simulation of over-delegation disaster.
  • Chapter 14: Clean Agents — the methodology for Supervised Autonomy.

Delegation is not abdication