Microsoft and Shanghai Jiao Tong University open-source Argus, a 1,548-hour autonomous research system

From "able to execute" to "able to steer," what is still missing for long-term agents?

Today, as agents rapidly evolve, Harness has enabled large models to interact with the real world by calling tools, modifying code, and running experiments. However, when tasks extend from tens of minutes to several days, the system still requires a person to remain at the screen for extended periods to determine the next steps.

The key bottleneck isn't just insufficient model capability, but that existing agents have largely automated the "execution" without truly automating the "navigation" above execution. Harness gives models the ability to act, while humans provide the decision-making. Once humans are removed, projects come to a halt. Additionally, without dense reward feedback, agents' execution and responses become sluggish and clumsy.

To address this issue, Microsoft, Shanghai Jiao Tong University, and other institutions open-sourced Argus, a general agent reasoning runtime designed for long-term research tasks, along with a technical report. The system enables agents to extend across multiple domains and conduct continuous research over several days without dense standard feedback, through design principles such as evidence-driven reasoning, self-evolution, multi-agent collaboration, and decoupling of core and domain-specific components.

The report covers 27 campaigns and 1,548 hours of wall-clock time, with an average of one human intervention requested every 40.7 hours; the report’s workload utilization ranges from 95.1% to 98.7%. This Argus delivery is not just a high-scoring benchmark, but a comprehensive set of production-ready outcomes. The team has delivered results across diverse domains including AI4AI, GPU kernels, model training, AI4Science, chip design, AI4Math, and AI4Systems, demonstrating Argus’s versatility and genuine research-grade intelligence.

Figure 1: Overview of Argus runtime, capability benchmarks, and delivery outcomes.

• Paper Title: Argus: Who Drives the Harness for Days?
• Paper: https://arxiv.org/abs/2608.05144
• Code: https://github.com/lbx154/Argus
• Project homepage: https://argusbot.cn/
• Project outcome open-source repository: https://github.com/Argus-AiTeam
• Project Math Problem-Solving Live Stream: https://open.argusbot.cn/#counterexample-live

Over the past two years, the primary advancement in Agent engineering has centered on the Harness: using the analogy of autonomous driving’s FSD, the model is like the engine, and the Harness is like the transmission—but the person sitting in front of the screen still determines where the vehicle goes. In short tasks, like automatic parking, the task itself provides clear feedback; but once tasks extend over several days, research problems often lack stable rewards and clearly defined goals from the outset. This exposes the true bottleneck in current Agents: they can execute, but they still cannot continuously judge under uncertainty.

Figure 2: Argus shifts the human role from the driver’s seat, constantly pushing forward, to the passenger seat by enabling autonomous research through auto-research.

Argus refers to this previously non-automated position above Harness as the Driver. Its role is to maintain control based on evidence, even when plans conflict with reality. The goals written at the start of the research are merely initial assumptions made during the stage of minimal information; if an Agent rigidly pursues preset endpoints in a Goal-Driven manner—even squandering Tokens on impossible objectives—it leads to goal rigidity. Evidence-Driven, by contrast, reverses the control logic: the next step is determined by existing evidence, not by the original assumptions of the plan. All outcomes generated during operation serve as valid evidence capable of altering the course; the system is fundamentally driven by evidence. Specifically, the Driver must repeatedly answer the following four questions based on current evidence:

Has this task been completed, and is the quality sufficient?

2. Based on the current evidence, what is the most worthwhile next step?

3. How should the lessons learned from this round change the system's future behavior?

4. Do the remaining issues involve decisions that only humans can make?

Argus organizes long-term projects into persistent Campaigns, which are then broken down into a series of well-defined Missions. Its fundamental feedback loop is: Manager → Planner → Engineer ⇄ Reviewer → Manager:

Referring to OpenAI Codex’s /goal mode clarifies Argus’s design orientation. /goal extends a single-cycle agent loop into a persistent loop with a goal state; Argus organizes research projects into long-running executions involving multiple roles, multiple harnesses, and accumulatable knowledge. The former answers, “How do we keep the agent from stopping?” while the latter answers, “Once the agent doesn’t stop, how do we ensure it works effectively?” This reflects the structural difference at the runtime level between a Goal-Driven and an Evidence-Driven approach.

1. The Manager controls stage transitions, process approvals, and global strategies;

2. The Planner outlines specific tasks based on current evidence;

3. Engineer accesses the real codebase, conducts experiments, and executes;

4. The reviewer independently examines the artifacts, assesses their innovation and effectiveness, and reports to the Manager or returns them to the Engineer.

The focus here is not on the multiple roles themselves, but on decomposing the context—multiple roles collaborate to prevent the agent from becoming a simple local hill-climbing algorithm. Each role has its own unique context but shares a common workspace, ensuring that planning, execution, and verification are not all delegated to a single agent, thereby improving token efficiency. This is precisely why, within a single Argus task, you can simultaneously enable Pi, Codex, and Claude Code. DeepSeek Harness Different harnesses ensure high customization.

The system saves a complete set of artifacts; only those that meet the evidence threshold are added to the Wiki and Skill and made available for reuse in subsequent tasks under Project, Vertical, or Global scopes.

Meanwhile, Core remains decoupled from Vertical: Core handles role permissions, evidence submission, and human boundaries, while Vertical, defined by domain experts, determines what constitutes valid evidence and the associated human skills and knowledge for tasks involving mathematics, GPUs, materials, and more; Vertical significantly improves the quality of research task delivery and provides an interface for the co-evolution of human experts and agents, as well as customized workflows.

Figure 3: Argus's long-term operation mechanism. Four roles cycle around a persistent state, and campaigns can advance or roll back across eight research stages.

What truly determines the viability of long-range agents is whether time can be converted into tangible research outcomes. Less than a month after open-sourcing Argus, we have already made outstanding contributions in infrastructure, materials, chips, mathematics, and AI systems research. Additionally, we are the first team to publish complete end-to-end logs of mathematical conjecture solving, opening all runtime session trajectories. Below is a summary of Argus’s achievements since open-sourcing:

Figure 4: Representative research outcomes, key metrics, and acceptance criteria of Argus

As shown in the table above, Argus first transformed continuous operation into verifiable, tangible deliverables within AI4System & Infra, achieving clear engineering outcomes that marked the first step from automated execution to autonomous research. Subsequently, the scope expanded from AI infrastructure to AI4Science, AI4Hardware, and AI4Math, with evaluation criteria shifting from “whether predefined tasks were completed” to “whether unresolved real-world research questions could be explored.” This advanced automated research from automated delivery to automated exploration. Building on this, Argus further extended its approach in AI4AI from isolated results to end-to-end research workflows, continuously driving progress through evidence—eliminating the need for constant human oversight. Thus, a progressive pathway has emerged, moving from tangible delivery, to cross-domain exploration, and ultimately to fully autonomous research.

Figure 5: Argus’s deliverables throughout the full paper production process.

All of the above results were achieved within one month. Connecting these achievements, Argus forms a progressively deeper value chain: the scope of automation expands from single executions to evidence-driven research advancement, significantly accelerating research progress—this is the most direct answer to the question, “Who drives the agent after humans step away from the screen?”

Long-range intelligence transforms tokens into intelligence, using that intelligence to continuously advance a research project and generate real-world value. Argus organizes previously disconnected model calls into a continuously operating research system, directing progress with evidence-driven decision-making and evolving experience into capability through self-improvement.

Argus doesn't just show a longer-running agent—it represents a new way of organizing research: Harness determines how models act, while Driver decides how the system keeps moving forward, freeing research speed from the constraints of human work and leisure cycles. This may signal the dawn of an accelerated era of research. The screen can turn off, but the research continues.

Author bio

Argus is led by researchers from Microsoft and Shanghai Jiao Tong University, with contributions from researchers at multiple universities.

This article is from the WeChat public account "Machine Heart".