Argus Agentic Runtime for Long Tasks
Listen to the summary
Uses a voice available on your device
Audio options
On this page 4 sections
Related concepts 3 concepts
Key Takeaways
- Argus achieves a 76.8% success rate on AARRI-Bench research tasks.
- The system outperforms Direct Copilot on SWE-Bench Pro, reaching approximately 78% success compared to 59%.
- Mature system waves demonstrate 21% fewer input tokens and 15% less active workflow time per task than startup waves.
- The runtime provides a framework for persistent memory and role-based coordination across long-horizon research missions.
Summary & Methodology Analysis
Argus operates as a persistent agentic runtime that maintains a durable project state across sessions to handle complex research tasks. By assigning model-driven roles such as Manager, Planner, Engineer, and Reviewer, the system organizes work into bounded missions. This architecture uses a report-level contract model to refine operational objectives while maintaining stable intent. The runtime decouples its self-evolution, which updates persistent state and control policies, from underlying model weight updates, ensuring that skills and memories are committed only after passing verification gates.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What is the primary goal of Argus?
Argus is a runtime system designed to manage long-horizon research tasks where initial objectives are often unclear.
Q2. How does Argus perform on standard benchmarks?
It reaches approximately 78% on SWE-Bench Pro compared to 59% for Direct Copilot, and it scores 76.8% on AARRI-Bench research tasks.
Q3. Does the system improve with use?
Yes, mature system waves require 21% fewer input tokens and 15% less active workflow time per task than initial startup waves.
Q4. What is the role of human guidance in Argus?
The system utilizes user-guided pivots to help manage objective refinement, though these are not yet publicly evaluated prospectively.
Q5. Does the system completely prevent goal drift?
No, while the use of evidence and explicit authority reduces goal drift, it does not eliminate it entirely.
Q6. Are the system's performance gains based on model weight updates?
No, the runtime performance improvements are achieved through persistent state and control policy updates, rather than weight-level fine-tuning or training.
Q7. How does the research compare startup versus mature performance?
The comparison is based on observational data rather than a controlled causal ablation.
Q8. What benchmarks were used to validate the system?
The system was validated on SWE-Bench Pro and AARRI-Bench.
Q9. How does the system handle model-specific benchmarks?
Argus shows competitive performance on GPU-kernel and language-model-training results, alongside a 28.0 gap on mathematical data synthesis.