Researchers from Tsinghua University, Zhipu, Shanghai Jiao Tong University, the University of Warwick, and Kuaishou have introduced HarnessPro—a method that enables artificial intelligence to improve the environment in which it operates. This environment is a framework that defines prompts, tools, budgets, and the sequence of agent tasks.
In a scientific paper published on arXiv on October 19th, involving Zhipu's lead scientist Tan Ze, a second goal for self-improvement was added: increasing performance while simultaneously reducing operational costs.
This project continues Zhipu's work from September, when information was presented about the GLM-5.3-based Infrastructure Agent, which helped create and configure the inference infrastructure for GLM-5.3-Flash, increasing end-to-end throughput by approximately three times compared to the initial level in less than two weeks.
How HarnessPro Works
HarnessPro starts with an existing configuration, collecting data on tool calls, environmental feedback, and token usage. Then, a trajectory analysis agent studies several runs of the same task to identify failure causes, as well as losses such as redundant searching, aimless exploration, and excessive checks.
After this, a planning agent develops three types of changes, prioritizing economy, balance, or performance. A framework editing agent implements each change in a separate workspace, which is tracked using Git for auditing and rollback capabilities.
To evaluate progress without using ground truth answers or official tests, a test generation agent creates tests based on the task description and early runs, after which these tests are fixed for each candidate. This approach significantly increased the failure detection rate from 41.7 percent to 83.3 percent in early experiments, compared to the log reading model query method for determining task success.
Candidates that perform worse than the original framework, or that lag behind another version in terms of cost and success, are excluded. Lite, Medium, and Heavy variants proceed to the next testing stage.
Testing Results
The team conducted testing on Terminal-Bench 2, SWE-Bench Pro, and ProgramBench using GPT-5.5 with Codex. Compared to the earlier RHO method, the Heavy variant increased the average pass rate by 2.0 percentage points and reduced the average task cost by 26.1 percent.
On Terminal-Bench 2, the pass rate rose from 76.3 percent to 80.8 percent, and the task cost decreased from $1.09 to $0.72. On SWE-Bench Pro, the Medium variant showed a 79.0 percent pass rate versus 80.0 percent for RHO, with the cost being $1.26 per task instead of $2.68.
The main efficiency gain was achieved by aligning reasoning efforts and tool budgets with each task, which prompted the agent to stop exploring, as well as by compressing repetitive logs and instructions. The financial figures provided only account for task execution, excluding the preceding search and evaluation processes.
