Braksa separates the agent’s proposal from its impact, using deterministic engineering to control cost and reveal model reliability limits

Researcher Stefan Greydor presents in a new paper on the arXiv platform an open-source framework for software-agent engineering called Praxa. The study begins with a fundamental engineering problem that developers of large-language-model-based AI systems face: the common conflation between an agent merely proposing an operational step and the agent actually possessing the programming authority to execute it, sending it, verifying its impact in the external environment, and promoting it to production servers. The new system seeks to make these five cases explicit and fully independent through mandatory auditing of input requests, mediated execution, reverse reading for verification, and sequential settlement operations that prevent the agent from making unobserved changes.

The study relied on four rigorous experimental tracks to test the software architecture and delineate its objective limits. In the first track, the internal audit of the approved code branch showed successful passage of 1,027 unit tests, plus 89 tests on the “Worker D” engine, with a review covering 363 source files, and the researcher explicitly acknowledged the absence of full recordings of individual tests and the lack of an independently reproducible experiment so far. In the second track, the author tested the reliability layer through an exploratory experiment on the “Terminal Bench Core” benchmark, which included 12 selected tasks, and the security system matched the baseline by achieving only 17 successful attempts out of a strict total of 36 attempts for each. More importantly, the trustworthiness layer consumed additional inputs by 37.49 percent and produced output symbols higher by 50.73 percent, leading the researcher to acknowledge that the experiment does not support any performance advantage.

In the refined comparison path for process coordination after bug fixing, the experimental version and the baseline completed 180 attempts with matching measurement precision and fully recovered from software crashes without recording any violation of the protected boundaries.The real gain in this experiment is a reduction in token consumption by 37.11 percent and an estimated endpoint cost decrease of 33.84 percent, achieved by cutting the steps by 11.63 percent.Despite this saving, the researcher notes that this reduction does not necessarily prove an improvement in overall quality or response time in live operating environments, and that examining review paths, accounting for retrieval, and tool validity did not show any tangible improvement in final production results.

ThisThis rigorous engineering analysis demands a direct review of the development teams, banks, and tech institutions in the Gulf, Egypt, and the Levant that are racing to deploy automation agents and delegate decision-making tasks to them. The operational lesson here is that separating proposal authority from implementation impact becomes essential to protect the digital architecture from phantom or uncontrolled actions. Moreover, adopting uncalculated security layers may increase cloud computing costs by more than half due to token output inflation without a commensurate improvement in delivery rates, which requires close monitoring of inference budgets.

The paper concludes that Braksa’s proven contribution lies in providing an architecture that makes the transition from delegation to external impact measurable and experimentally verifiable.Current evidence does not demonstrate security immunity against hostile attacks, nor complete safety in production environments, nor any general domain-specific superiority or capacity for iterative self-improvement.Your practical step as a systems engineer is to implement reverse-reading mechanisms that require the agent to verify the result from independent system logs, rather than relying on its own assertion that the task is complete.