Xenon
An agent harness for AI coding agents built around treating tool output as evidence and language-model output as an unverified claim, with execution sandboxing, seven interchangeable reasoning strategies, and a published, reproducible SWE-bench Lite score.
Xenon frames itself as infrastructure for running a coding agent trustworthily rather than another coding assistant on its own. Its verification layer treats a tool's output as evidence and a language model's own statements as an unverified claim, intended to interrupt hallucinations before they propagate; its constraint layer enforces path fencing, command-injection interception and permission gates, with all file side effects routed through one convergence point. Seven interchangeable reasoning strategies are available, including ReAct, plan-execute and reflection, registered through a pluggable interface.
On SWE-bench Lite, the project reports a 40.0 percent instance-level pass rate on the same model, an improvement of 6.7 percentage points over a baseline, and a 45.8 percent pass rate across a five-engine by thirty-instance matrix, alongside a 97.35 percent cache hit rate that it reports cuts token cost by 93 percent, measured through the official SWE-bench Docker harness. Runtime-resilience features include checkpointing that resumes a generation after a network interruption or model switch without duplicating content already produced, loop detection shared across every reasoning engine, and detection of evasive or padded non-answers.
It runs as a command-line tool or as a library embedded in an agent evaluation pipeline, requires Python 3.10 or later, and is documented in Chinese and English.