Microsoft's Argus: Evidence-Driven Agent Runs 1,548 Hours of Autonomous Research

From Goal-Driven to Evidence-Driven: Who Decides the Agent's Next Step?

Current agents excel at short tasks where clear feedback exists, but for multi-day research they still require a human to continuously steer. The bottleneck is not model capability alone; existing agents automate execution but not the higher-level "driving" that decides what to do next when rewards are sparse and goals are ill-defined. Argus introduces an Evidence-Driven Driver that replaces rigid goal-chasing with a loop where every result becomes evidence for the next decision. The Driver repeatedly answers four questions: (1) Is the work complete and of sufficient quality? (2) Given current evidence, what is the most valuable next step? (3) How should this round's experience change future system behavior? (4) Does the remaining problem require a human-only decision?

Building a General-Purpose Customizable Long-Horizon Runtime

Argus organizes long-term projects into persistent Campaigns composed of well-bounded Missions . Its core loop is Manager → Planner → Engineer ⇄ Reviewer → Manager :

Manager handles phase transitions, approvals, and global strategy.

Planner creates concrete tasks based on current evidence.

Engineer works in real codebases, runs experiments, and executes code.

Reviewer independently audits artifacts for correctness and novelty, reporting to Manager or sending back to Engineer.

Separating contexts across roles prevents a single agent from becoming a local hill-climber and improves token efficiency. This design also enables multi-harness support : within one task, Argus can simultaneously invoke Pi, OpenAI Codex, Claude Code, DeepSeek Harness, etc. Only experiences that pass an evidence threshold are distilled into a shared Wiki (knowledge) and Skill (reusable procedures), scoped at Project, Vertical, or Global levels for later reuse.

A key architectural principle is Core–Vertical decoupling . The Core manages role permissions, evidence submission, and human-in-the-loop boundaries. Verticals are domain-specific modules (mathematics, GPU kernels, materials, etc.) that define what counts as valid evidence and provide expert skills. This allows domain experts to customize workflows while the Core remains stable.

1,548 Hours Later: What Argus Delivered

The technical report covers 27 Campaigns totaling 1,548 wall-clock hours . The system requested human intervention on average every 40.7 hours and achieved a duty cycle of 95.1%–98.7% . Results span multiple domains:

AI4System & Infra : Turned continuous operation into verifiable engineering deliveries.

AI4Science, AI4Hardware, AI4Math : Shifted evaluation from "task completion" to "exploring open research questions."

AI4AI : Extended from point results to full research pipelines, driven continuously by evidence.

Notably, the team claims to be the first to publish complete end-to-end logs for mathematical conjecture solving (including all session traces). All artifacts are open: paper ( https://arxiv.org/abs/2608.05144), code ( https://github.com/lbx154/Argus), project homepage ( https://argusbot.cn/), results repository ( https://github.com/Argus-AiTeam), and a live math-solving demo ( https://open.argusbot.cn/).

Conclusion: Research Continues After the Screen Goes Dark

Argus demonstrates a new research organization paradigm: Harness solves how models act; Driver solves how the system keeps moving forward. By converting tokens into sustained intelligent progress, it decouples research velocity from human working hours. The screen can go dark, but the research continues.