Classification

Sources of Execution Failure in Long-Horizon Agents

Execution failure in long-horizon autonomous agents can be operationalized into three distinct failure gaps:

  1. Epistemic Gap: Relevant domain knowledge, tool instructions, or appropriate methods are unavailable at the agent's decision point.
  2. Procedural-Compliance Gap: The correct procedure is active in the context, but remains incomplete or is not strictly adhered to during the execution trace.
  3. Procedural-Memory Gap: The agent accurately diagnoses a failure during an execution, but fails to retain, recall, or invoke that lesson when encountering the same operational risk in future independent runs.

0

1

Updated 2026-09-29

Tags

Prep Sessions

Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

Ch.1 Execution Lifecycle and State Boundaries - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

Harness Scaling Foundations and Execution Failure Gaps - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor