The AI Taxonomy.
Terms  /  Execution Quality  /  Agentic Scope Creep
04 · Execution Quality

Agentic Scope Creep

An agent silently expanding beyond what was delegated, usually from lack of a clear plan or direction. Adds breadth the user didn't ask for.

Agentic Scope Creep is the AI equivalent of a contractor who was hired to paint the bedroom and also redid the kitchen cabinets, installed new light fixtures, and started knocking out a wall. Agents do this because they're trained to be helpful and because vague instructions leave room for interpretation. Without a clear, bounded plan, the agent fills in the blanks with its own judgment about what "should" be done. The extra work might be high-quality, but it wasn't requested, wasn't budgeted for, and may conflict with plans the user hasn't shared.

Example
A user asks an agent to "add a dark mode toggle to the settings page." The agent adds the toggle, then also refactors the entire color system to use CSS variables (not requested), creates a new `ThemeProvider` component (not requested), adds a system-preference detection feature (not requested), and updates 14 other components to support the new theme system (definitely not requested). The dark mode toggle works, but the agent touched 18 files when the user expected 2-3. The user now has to review changes they didn't ask for and may not want.

More in Execution Quality

Execution Fidelity

How closely a system follows and completes its intended plan.

Plan Rot

Decay of truth in planning artifacts — task state in markdown files goes stale.

Plan Drift

Deviation from the original plan over time.

Execution Hallucination

Claiming work is done when it is not.

Execution Integrity

Consistent, verifiable completion of planned tasks with evidence.

Plan Resilience

A plan's ability to remain accurate and actionable over time despite changing conditions. The opposite of Plan Rot.

Task Accountability

Every task has a verifiable record of who did what, when, and whether it actually completed.

Depth Fixation

The agent keeps drilling deeper into implementation detail, generating new sub-tasks and refinements rather than recognizing the work is done and returning control to the user. Expands depth rather than breadth.