The project's logic

QbitAI presents Agnes Harness as an open-source system, distributed under the Apache 2.0 license and also intended for commercial use, which has reached 2.1k stars on GitHub. The starting point indicated by the source is a shift in the economics: if software development once cost 90 and adaptation for each customer cost 10, with AI development falls to 1 while adaptation remains at 10. According to the source, it is on this latter item that the difficulty concentrates for those bringing agents into business and industrial processes.

The article argues that many agents are slow and costly not because large models are weak, but because they are loaded with tasks that could be separated. Hence the choice to split the work between a large model and a specialized decision model called Agnes Sys1.0, with 9 billion parameters.

The figures cited by the source

QbitAI reports several comparisons contained in the project materials. In a simulation of a household water leak, evaluating the next step would require an average of 7.6 seconds for a large model and 0.53 seconds for Agnes Sys1.0. In a test on a baggage room, involving weighing, vertical positioning and temporary storage, the cost per individual decision would be about $0.00247 for the online model and about $0.000091 for Sys1.0, over 100 judgments each.

The combination of a large model and Sys1.0 would also have reduced total task time by 32.9% compared with using only the large model at every step. The source also cites scores on two benchmarks: 87.07 on DeepPlanning, dedicated to planning complex tasks, and a 40.69% success rate on BigCodeBench-Hard V1, which concerns writing complex code. These are figures reported by the project and reproduced by the source, not an independent verification.

Temporary learning and shared memory

The system envisages that Sys1.0 adapts to the imminent task through exercises with known answers and modifies only a limited portion of its parameters, restoring them at the end. The stated reason is to prevent knowledge accumulated on different devices from ultimately interfering with one another. At a broader level, Agnes Harness re-examines execution to identify recurring errors and slow steps, then intervenes on prompts, memory, skills and subagents; corrections deemed valid are collected in reusable, versioned manuals.

The source specifies that the methods proposed by the AI must pass simulation tests before entering the manuals. The project also includes a plugin mechanism for tools, model adapters, workflows and interfaces, with version management: new sessions use the updated version, while tasks already underway remain on the previous one. For installation, Node.js 24.10 or higher and pnpm 10.34.5 are indicated; using cloud model APIs would not require an additional GPU.

The crux of physical integration and the limits

To connect software and real devices, Agnes Harness proposes MHS (Model Hardware Standard), a uniform communication specification for equipment. The source describes a test in which the agent coordinates search devices, a robot and a balance chamber to bring an astronaut back to a base. Some aspects nevertheless remain to be clarified: the material does not indicate which commercial devices are already compatible with MHS, nor does it offer external assessments of the results. The source also reports negative experiences of operators, including a case of thousands of failed attempts and high costs, used as context for the problem. According to the article, the real test is transforming what is learned in one project into reusable capability for the next, so as to reduce the adaptation cost that AI has not eliminated.

Source and attribution: this original summary was prepared by AICronaca based on the report by QbitAI ↗. The statements are attributed to the source; any informational limitations: The performance, costs and scores are those declared by the project and reproduced by QbitAI; no independent verifications are available. No customers, compatible commercial devices or release dates are indicated..