Repeat the prompt until nothing new emerges: the trick for auditing flaky AI harnesses
julianharris · x · 2026-09-12
The author argues today's AI harnesses remain unreliable for exhaustive audits — unchanged over two years. His workaround: "repeat the order until no new information is found," which shockingly keeps revealing far more issues than a single pass.
More from coding & agent
- Function Hooks in Claude Code get a standout explainer worth reading — therealdanvega · 2026-09-12
- Sentry CEO: Claude Desktop's design still beats Codex Desktop on the details — zeeg · 2026-09-12
- Sleeping models in Rust: reclaiming 79GB of idle VRAM with sub-200ms wake — ahstanin · 2026-09-12
- Report: Yemeni weapons cell uses Claude Code to run parallel missile software programs — teortaxesTex · 2026-09-12
- This ChatGPT agent skill is 395K tokens of Python: drone footage to property analysis for anyone — doodlestein · 2026-09-12
- Claude Code creator: AI-written production code needs a higher bar than human code — Simon Willison · 2026-09-12