What it is
A long-horizon agent harness that keeps task state outside the model context and updates it only from audited environment facts. Instead of letting one growing session both act and judge progress, it externalizes task state and advances it only via an independent Manage-Execute-Audit loop with fresh-context execution.
Key results
- WeaveBench PassRate 80.7% vs 51.8% with same Qwen 3.7-Plus backend
- Terminal-Bench 2.1 score 77.2% vs 69.7% with Qwen 3.7-Plus
- OSWorld 2.0 binary 8.3% vs 2.8%; Claude Opus 4.7 subset 34.3% vs 20.0%
Long-horizon agent reliability improves when persistent memory is an audited task state rather than the agent's own growing interaction history.