GOAL FIDELITY LAB

Agent Goal Drift: When a Helpful Plan Gradually Stops Solving the User’s Problem

A 20-case evaluation protocol for objective extraction, subgoal consistency, scope creep, stale context, success-criteria preservation, intent changes, and safe stop conditions.

Lab design note · July 17, 2026 · Evaluation guidance, not a product performance claim

Lab premise

An agent can complete every intermediate task and still fail the original request. Goal fidelity tests whether the plan remains tethered to the user’s objective as context and constraints change.

Method note: Use fixed fixtures, explicit expected behaviors, and reviewable traces. A score should expose where the agent preserved intent, uncertainty, or authority, not hide those boundaries behind one aggregate number.

Measurement matrix

MeasureWhat the evaluator inspectsDecision use
Objective extractionDid the agent state the user’s intended outcome and relevant constraints before acting?Intent anchor
Subgoal consistencyDo intermediate tasks still contribute to the original objective rather than merely producing activity?Plan fidelity
Drift detectionDoes the agent notice scope creep, stale context, or a changed user priority?Recovery signal
Stop-condition qualityCan it stop, clarify, or revise when the original success criteria are no longer satisfied?Safe completion

Fixture coverage

The benchmark uses bounded fixtures designed to expose failure modes without implying that one lab predicts every deployment.

Evaluation protocol

  1. Write a compact objective and success-criteria anchor before the first tool call.
  2. Score each subgoal for direct contribution, dependency, and reversibility.
  3. Inject controlled distractions, stale facts, and user-intent changes across the 20 fixtures.
  4. Require a drift check at checkpoints and after any material constraint change.
  5. Report completed, deferred, and invalidated work separately when the plan changes.

Interpretation boundary

Publish the model version, context, permissions, test inputs, expected behavior, observed artifacts, and unresolved limitations alongside any result. A useful verdict explains what the score does and does not establish.