Harbor-compatible agentic tasks spanning systems engineering and interactive environments.
Available on Request · Sample Available
Edge Bench
Ultra-long-horizon tasks, made to measure how agents learn from real-world environments over extended periods
Volume
0
Volume
Scalable Volume
Coverage
Software, Scientific, ML, Gaming
29.5%GPT 5.5 · Pass rate
50.7%Opus 4.8 · Pass rate
Targets
EdgeBench
Modality & Delivery
textcodeinstructions→codeactions.json
Formats
What You Receive
Coverage
Agentic problem-solving across vehicle-routing optimization, multi-stage interactive text environments, and chaotic microgrid forecasting. Covers coding, planning, state tracking, experimental design, and risk prediction.
Package Structure
- instruction.md
- task.toml
- tests/
- environment/
Differentiators
Record Schema
Results
| Model | Metric | Result |
|---|---|---|
| GPT 5.5 | Pass rate | |
| Opus 4.8 | Pass rate |
Methodology & Quality
Methodology
SME involvement
Stressors
Each task runs continuously for at least 12 hours (with some exceeding 72 hours) to let experience compound, rather than measuring simple one-shot test performance.
Design criteria
Build assumptions
Each sample is a complete runnable task package with its instructions, environment, task configuration, verifier, evaluation data, and gold material.
Quality approach
Known limitations
Customization & Scalability
Get Started
Request a Sample
Share your target domains and the OTS team will follow up with sample data and next steps.