Researchers present Argus, a runtime system designed to run long, multi-step research and engineering tasks while keeping a clear record of why it changed course. Argus separates a stable user intent from the current operational objective, constraints, and verification rules. It lets an automated workflow keep working when evidence supports the current approach and safely change direction when measurements show a route is failing — but only when the change is backed by evidence and checked by a separate role.

Argus divides work into bounded missions that run against a durable project state. Four roles govern the process: a Manager anchors the campaign and authorizes major changes; a Planner chooses the next unit of work; an Engineer carries out and tests the work; and a Reviewer inspects results and issues formal completion decisions. Candidate memories, skills, procedures, verifiers, routing choices, and rejected routes are written into persistent state only after role-owned review and, when available, task-native verification. The system keeps model weights fixed, so “self-evolution” happens by changing the runtime state and control policy rather than retraining the model.

The paper gives concrete results from a broad benchmark suite and from real campaigns. In one set of software-engineering tests (SWE-Bench Pro), Argus reached about 78% success versus 59% for a Direct Copilot baseline while using 1.41× the total tokens. As the runtime admitted verified changes and matured, later waves used 21% fewer solve-input tokens and 15% less active workflow time per task than early waves. The authors also report 34 cases where verifiers recovered a failing route and 22 strict reviewer interventions that rescued stalled work. Across other arenas, Argus scored 76.8% on AARRI-Bench and showed a 28.0-point advantage on a mathematical data synthesis task. Outside benchmarks, an optimized RWKV6 kernel was merged upstream, a multi-day math campaign retained a falsified route and six proof-backed frontier updates, and six paper-pipelines ran 254 missions with 16 stage rollbacks.