Sources of Execution Failure in Long-Horizon Agents
Execution failure in long-horizon autonomous agents can be operationalized into three distinct failure gaps:
- Epistemic Gap: Relevant domain knowledge, tool instructions, or appropriate methods are unavailable at the agent's decision point.
- Procedural-Compliance Gap: The correct procedure is active in the context, but remains incomplete or is not strictly adhered to during the execution trace.
- 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
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
Learn After
An epistemic gap describes a scenario where an agent has access to complete tool documentation in its active context but deliberately disregards it during execution.
Contrast the operational scope of a procedural-compliance gap with that of a procedural-memory gap across agent executions.
Match each operational failure scenario to the corresponding execution failure gap.
An epistemic gap arises when required domain knowledge, tool instructions, or appropriate methods are missing at the agent's ___ point.
Arrange the events in chronological order to illustrate how a procedural-memory gap develops across agent executions.