Argus: Evidence-Driven Runtime for Autonomous Long-Horizon AI Research

Introduction: The Missing Driver in Long-Horizon Agents

Current agent frameworks like Harness give models the ability to act — call tools, modify code, run experiments — but they still require a human to continuously decide what to do next when tasks stretch from minutes to days. The bottleneck is not only model capability; it is that existing agents automate execution but not the higher-level "driving" function that judges progress, pivots based on evidence, and decides when human input is truly needed. Without dense reward signals, agents become sluggish and waste tokens chasing rigid, pre-defined goals.

01 From Goal-Driven to Evidence-Driven: Who Decides the Next Step?

Argus introduces a Driver layer above the Harness. Instead of blindly pursuing an initial goal (Goal-Driven), the Driver constantly re-evaluates direction based on accumulated evidence (Evidence-Driven). The initial goal is merely a low-information hypothesis; every runtime result — success, failure, surprise — becomes evidence that can change the route. The Driver operates by repeatedly answering four questions:

Is this work complete and of sufficient quality?

Given current evidence, what is the most valuable next step?

How should this round's experience change future system behavior?

Do remaining problems require decisions only a human can make?

This shifts the human from constant driver to occasional co-pilot, intervening only when the system flags a genuine need.

02 A General-Purpose, Customizable Long-Horizon Runtime

Campaigns, Missions, and the Four-Role Loop

Argus organizes long-term projects into persistent Campaigns , decomposed into well-bounded Missions . The core loop runs: Manager → Planner → Engineer ⇄ Reviewer → Manager .

Manager controls phase transitions, approves workflows, sets global strategy.

Planner creates concrete tasks grounded in current evidence.

Engineer executes in real codebases, runs experiments, produces artifacts.

Reviewer independently audits artifacts for correctness and novelty, then reports to Manager or sends back to Engineer.

Unlike simply extending a single agent's loop (e.g., OpenAI Codex /goal mode), Argus separates contexts across roles. Each role has its own context window but shares a workspace, preventing the agent from collapsing into a local hill-climber and improving token efficiency. This design also allows mixing multiple harnesses — Pi, Codex, Claude Code, DeepSeek Harness — within one Campaign.

Knowledge Persistence: Wiki and Skill

Only experiences that pass an evidence threshold are distilled into a Wiki (project knowledge) and Skill (reusable procedures), scoped at Project, Vertical (domain), or Global levels for future reuse.

Core–Vertical Decoupling

The Core handles role permissions, evidence submission, and human-in-the-loop boundaries. The Vertical layer, defined by domain experts, specifies what counts as valid evidence in mathematics, GPU kernels, materials science, etc., and provides domain-specific skills and workflows. This decoupling lets researchers customize the runtime without touching the core engine.

03 Results After 1,548 Hours: A Complete Production-Grade Portfolio

The technical report covers 27 Campaigns totalling 1,548 wall-clock hours . Key metrics:

Average proactive human intervention request every 40.7 hours .

System duty cycle (active work time) between 95.1% and 98.7% .

Results span seven domains, demonstrating generality and research-grade intelligence:

AI4System & Infra — concrete engineering deliverables turning continuous operation into verifiable deliveries.

AI4Science, AI4Hardware, AI4Math — exploration of open research questions, shifting evaluation from "task completion" to "exploration of unresolved problems."

AI4AI — full research pipelines from hypothesis to paper, driven continuously by evidence without human babysitting.

The team also released the first complete end-to-end logs for mathematical conjecture solving, including all session trajectories (live view: 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. Research velocity is no longer bound by human working hours. The screen can turn off; the research keeps running.

References & Resources

Paper: https://arxiv.org/abs/2608.05144

Code: https://github.com/lbx154/Argus

Project homepage: https://argusbot.cn/

Artifacts repository: https://github.com/Argus-AiTeam

Math solving live logs: https://open.argusbot.cn/