Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning
Argus is a persistent, self-evolving runtime with Manager, Planner, Engineer, and Reviewer roles executing bounded missions over durable project state. It separates stable user intent from operational objectives, constraints, and verification criteria. Model weights remain fixed; self-evolution occurs through persistent runtime state and control policy. Across seven GPT-5.5 benchmark arenas, Argus achieves about 78% on SWE-Bench Pro versus 59% for Direct Copilot while using 1.41 times the aggregate tokens. After verification-gated self-evolution, mature SWE-Bench waves use 21% fewer solve-input tokens and 15% less active workflow time per task than startup waves, while recording 34 verifier recoveries and 22 strict review-loop rescues.
Argus introduces a persistent, self-evolving agentic runtime designed for long-horizon reasoning tasks. It employs four specialized roles—Manager, Planner, Engineer, and Reviewer—that operate over durable project state, separating stable user intent from operational objectives and constraints. The system evolves through runtime state and control policy without changing model weights. In evaluations using GPT-5.5, Argus reaches approximately 78% on SWE-Bench Pro, outperforming Direct Copilot's 59% at a cost of 1.41× the tokens. Self-evolution leads to efficiency gains: mature runs use 21% fewer solve-input tokens and 15% less active workflow time, with 34 verifier recoveries and 22 review-loop rescues.
Argus achieves self-evolution without fine-tuning by maintaining persistent runtime state and a control policy that governs role interactions, memory, skills, procedures, and routing decisions. Verification-gated evolution allows the system to learn from failures and successes, reducing token usage and workflow time over successive waves. The architecture's separation of user intent from operational objectives enables robust long-horizon planning and error recovery.
The 78% vs. 59% improvement on SWE-Bench Pro suggests that agentic runtimes with persistent state and role-based coordination can significantly boost software engineering task performance. The 1.41× token overhead indicates a trade-off between accuracy and cost, but the self-evolution efficiency gains (21% fewer tokens, 15% less time) may make such systems more viable for enterprise deployment over time.
For organizations building autonomous coding agents, Argus demonstrates a path to higher accuracy on complex tasks with manageable cost increases. The self-evolution capability could reduce long-term operational costs and improve reliability, making it attractive for continuous integration and automated software maintenance workflows.
Next signals to watch include: (1) whether the Argus runtime can be adapted to other foundation models beyond GPT-5.5; (2) real-world enterprise pilot results that validate the benchmark gains; (3) further reductions in token overhead through optimized control policies; and (4) integration with external tools and APIs to expand beyond software engineering tasks.