Argus
A persistent, multi-role runtime for long-horizon AI research and engineering, splitting planning, execution and read-only review across Manager, Planner, Engineer and Reviewer roles so work can pause, be checked and resume for hours unattended.
Argus is a runtime for long-running AI agent work that can plan, execute, verify, pause and continue across many model turns rather than a single conversation. It splits the work traditionally held by a human across four roles that are deliberately not allowed to do each other's jobs: a Manager that owns stage transitions and what lessons get kept, a Planner that sets the next task and the evidence it must produce, an Engineer that implements and researches, and a read-only Reviewer that judges correctness and evidence and can send work back as blocked.
It improves over time without retraining by admitting skills and source-linked findings scoped from a single project up to a global level, and ships 24 domain-specific extensions built on an unchanged core. Credentials, payments, irreversible actions and publication always stop for a human to approve.
It supports several backends including GitHub Copilot CLI and OpenAI Codex CLI, and is developed at microsoft/ArgusAgent, with a separate development-preview repository that periodically synchronises into the official one.