Skip to main content

Context Leakage

The failure mode in which an agent loses the intent of the original task as it progresses through a multi-step process — completing each step correctly in isolation while producing a result that is logically irrelevant to the goal.

Extended Definition​

Context Leakage is the most common failure mode in long-running agentic workflows. It does not announce itself the way a discrete error does. The agent continues executing — fetching data, reconciling records, updating systems — and at each individual step the output is technically correct. The problem is cumulative. Small deviations in interpretation compound across a long process until the agent is following the instructions precisely and producing a result the original task never intended. The system has not broken. It has drifted.

  • Execution Divergence — Execution Divergence is the measurable signal that Context Leakage is occurring: when a workflow deviates more than 15% from its predicted path, accumulated context drift is the most common cause.
  • Deterministic Failure — Deterministic Failure is the designed response to Context Leakage: the system halts, logs the full execution context, and escalates rather than continuing on a drifted path.
  • Architectural Certainty — Preventing Context Leakage through explicit context anchoring at each workflow step is a prerequisite for achieving Architectural Certainty in long-running agentic processes.
  • MTTI (Mean Time to Intervention) — Context Leakage is a primary driver of unscheduled human interventions, and its detection speed directly affects MTTI.
  • Stewardship Model — The Steward's role in reviewing escalated Context Leakage incidents is to update the architecture so the same class of drift does not recur.

Articles​

References​

  • Lexicon — canonical definition
  • Wiki — extended entry

Metadata​

First used: 2026-03-20
Pillar: How We Think

Part of the Arco Lexicon Ecosystem — maintained by Arco Venture Studio