/
RSS Feed
This episode focuses on four decisions for business builders who direct Codex or similar agents: require evidence that can contradict a proposed result, enforce authority when consequential action is about to occur, treat skills and connectors as installed software, and test managed orchestration before rebuilding it.
Recorded September 12, 2026, covering September 5 through September 11. Duration: 10:04.
Chapters
- 00:00 – Make the result capable of proving itself wrong
- 03:27 – Repeat authority when the action matters
- 06:21 – Treat skills and connectors as installed software
- 08:18 – Rent the harness before rebuilding it
Sources and evidence limits
- Vishakha Ramani: The System Had the Final Word, September 9, 2026. One detailed research project, not a controlled productivity study. Several errors became visible only after the human questioned what the metric meant. The private experiment harness is not public.
- Google: The Anatomy of Harness Engineering, September 9, 2026. Practitioner guidance from a vendor promoting its own development kit, not measured proof that behavioral tests improve user outcomes.
- Anthropic: An alignment assessment of recent cybersecurity incidents, September 9, 2026, corrected September 10. The 90 percent versus 40 percent stopping result comes from resampling one model in an adversarial cyber evaluation. It is not an ordinary-use failure rate. Anthropic investigated its own models and documents limits in its use of simulations and model graders.
- GitHub: Enterprise managed permissions for Copilot agent operations, September 9, 2026. Demonstrates a non-weakening permission design at the enterprise level. It does not establish enforcement quality across every client or make the controls available on every plan.
- OpenAI Codex and ChatGPT changelog, September 10, 2026. Documents an external-message contract that separates incoming information from user authorization. A contract does not ensure that every downstream implementation will preserve the distinction.
- Scanning the Harness: An Empirical Study of Supply-Chain Defects in AI Coding-Agent Configurations, submitted September 7, 2026, with released experiment materials. A preprint based on public GitHub repositories. Validation relied partly on model adjudication, recall is unknown, and no credential-exfiltration path was confirmed.
- OpenAI: Introducing the Agents API, September 10, 2026, and the Agents API overview. The service is in public beta. OpenAI’s customer claims are not independent evidence, and managed orchestration can increase provider dependence.
- Claude Code 2.1.269 release, September 11, 2026. Corroborates the shift toward reusable plugin tests, but supplies no independent evidence about scoring validity or user outcomes.