University of Michigan - Ann Arbor

Long-horizon agent execution failures are structured into three distinct operational gaps, each mapped to a specific control intervention in a stateful harness runtime:

1. **Epistemic Gap**: Relevant knowledge or methods are unavailable at the decision point; targeted by state-local context and entry hooks (`in_hook`) that reintroduce phase-specific guidance.
2. **Procedural-Compliance Gap**: An active procedure is left incomplete or not adhered to within a run; targeted by checked transitions, contract boundaries, and pre-commit checks (`before_transfer`) that block invalid handoffs.
3. **Procedural-Memory Gap**: Lessons diagnosed in earlier executions fail to be retained across runs; targeted by versioned practices and reusable runbook rules that accumulate procedural experience outside model weights.

Mapping Operational Gaps to Runtime Control Points

Which operational gap is characterized by relevant knowledge or methods being unavailable at an agent's decision point?

Which operational gap describes execution lessons failing to carry over across runs, and what runtime control points target it?

In a stateful harness runtime, pre-commit checks (`before_transfer`) and checked transitions are designed to block invalid handoffs when an active procedure is left incomplete.

Learn Before

Related