Researchers propose Praxa, an agent harness designed to make authority-to-effect transitions explicit and testable

Researchers introduced Praxa, an agent harness designed to make the transitions from proposal to verified effect explicit and testable in large-language-model (LLM) agents. The system models these transitions through deterministic admission, brokered execution, external read-back, reconciliation, and reviewed promotion, aiming to improve governance and auditability of agent actions.

The authors report four evidence lanes to support Praxa’s claims. First, an author-run repository-local audit at a pinned revision passed all 1,027 unit tests and 89 Workerd tests, instrumented all 363 expected source files, and met four coverage floors. However, raw per-test transcripts and independent reproduction were not available.

In a provider-backed Terminal-Bench Core 0.1.1 pilot across 12 curated tasks, both baseline and reliability-layer arms passed 17 of 36 strict trials. The reliability layer used 37.49% more input tokens and 50.73% more output tokens, so the pilot does not support claims of superiority.

A post-debug, two-order coordination-proxy development comparison showed baseline and a source-authored candidate each completed 180 of 180 trials with equal measured accuracy, full hermetic crash recovery, and zero protected violations. The candidate version used 37.11% fewer tokens, 33.84% lower estimated endpoint cost, and 11.63% fewer steps, but the authors note this does not establish improved quality, latency, or production behavior.

Deployed source and configuration evidence demonstrated bounded reflection, recall accounting, memory compilation, and tool-health paths, but no production outcome lift was observed. The paper explicitly states that current evidence does not establish adversarial security, production safety, general specialist superiority, autonomous recursive optimization, or user benefit.

The authors position Praxa’s contribution as an evidence-bound architecture that makes authority-to-effect transitions explicit and testable, rather than claiming immediate performance or safety advantages.