AI 代理與開發工具
Argus Enables Fixed-Weight Agents to Accumulate Experience, Raising SWE-Bench Pro Accuracy From 59% to About 78%
Argus breaks long-running tasks into constrained missions, with a Manager, Planner, Engineer, and Reviewer jointly maintaining persistent state. Its authors report improving program repair performance at 1.41× the token cost; an agent-generated RWKV6 kernel was also merged into an open-source project after external corrections.
Argus does not update model weights. Instead, it turns the agent execution layer into a state machine whose state can accumulate and be reviewed. The Manager preserves the user’s original intent, while the Planner, Engineer, and Reviewer work through missions with explicit budgets, objectives, and validation criteria. Memories, skills, validators, routing decisions, and even failed approaches are written to long-term state only after role-based review or task-native testing. This prevents a single error from directly contaminating subsequent work and allows the system to retain knowledge of approaches already ruled out across sessions.
The paper uses different native metrics across seven task categories. In its GPT-5.5-powered SWE-Bench Pro experiment, Argus achieved about 78% accuracy, compared with 59% for a control using Copilot directly, although Argus consumed 1.41× as many total tokens. As state accumulated, input tokens per problem in the mature stage were 21% lower than during startup, while active agent time fell by 15%. The Reviewer workflow recorded 34 validator rescues and 22 strict-review-loop rescues. These figures support the idea of making an upfront investment to enable later reuse, but they do not establish that costs will fall across all workloads.
More concrete external evidence comes from a TileLang RWKV6 kernel that the agent generated for Flash Linear Attention: forward-pass latency on a specified H100 workload fell from 0.199 ms to 0.168 ms. During maintainer review, a risk of exponential overflow on long sequences was identified, prompting a request to add block-local centering. The code was merged only after the agent made the correction and reran the tests. This demonstrates both the value and the limitations of validation gates: agents can produce work suitable for adoption, but domain experts are still needed to catch numerical issues that benchmarks do not expose. Engineering teams should next look for publication of the complete runtime, execution traces, and reproducible evaluations, and assess whether persistent memory causes errors to accumulate when switching tasks or models.