Same Prompt, Different Harness: Astra xHigh Yields Notably Different Results in Codex vs AGNT
NathanWilbanks_ · x · 2026-09-05
Developer Nathan Wilbanks ran the exact same prompt and seed image ("create this as a real interactive UI in html/three.js") through Astra xHigh in both Codex and AGNT, concluding that "the harness still matters" — the same model produces visibly different output depending on the execution framework. He posted a side-by-side comparison (left: Codex, right: AGNT) for others to judge.
Related event: Same Prompt, Different Results: Coding Harness Shapes Model Output Quality(3 posts)→
More from coding & agent
- Microsoft's FrontierTour: AI governance becomes everyone's problem as agents take action — donasarkar · 2026-10-23
- Grady Booch revives Weinberg's 1971 quip: software so buggy one woodpecker could destroy civilization — Grady_Booch · 2026-10-23
- One Agent Replaces Half an SDR Team: Full Lead-Research-to-Email Workflow — tibo_maker · 2026-10-23
- Why don't coding agent harnesses ship with their own sandbox? Only amp does — Aryvyo · 2026-10-23
- Dev builds a VST synth with Codex and MCP, lets the AI jam in his studio — Yamapama · 2026-10-23
- AgentsInTheCloud update: use your Claude subscription with built-in coding agents — lucasmeijer · 2026-10-23