Long-Horizon Harness
An evaluation framework for AI agents across extended tasks involving memory, context changes, interruptions, recovery points, and final artifact quality.
Short benchmarks miss the failure modes that appear in real work. Long-horizon evaluation tests whether agents can sustain coherent execution over time.
Problem Space
Agents may succeed on short tasks while failing across multi-step work involving changing context, partial progress, interruptions, and handoffs.
System Direction
The Long-Horizon Harness evaluates sustained execution through traces, checkpoints, evidence artifacts, and recovery-oriented review.
Public Capabilities
- 01Extended task evaluation
- 02Memory and context-change testing
- 03Interruption and recovery scenarios
- 04Artifact-quality review
- 05Trace-based performance analysis
The Long-Horizon Harness is described publicly as an evaluation direction. Internal tasks, scoring rubrics, traces, and datasets are not disclosed.
What Is Not Disclosed
Private implementation details, security-sensitive internals, and unreleased runtime architecture are intentionally not disclosed.