Running Agents in Prod for a Year: Auditability Trumps Model Capability
KimLikeJ · reddit · 2026-08-04
A developer who has been running AI agents against real production systems for a year shared core engineering insights. They pointed out that 'model capability' isn't the critical variable; how the system handles an agent being 'confidently wrong' is what matters.
- Hard lessons: Early on, they let an agent handle deployments without pre-review. This went fine until a database migration that looked correct in the diff broke a live site because tests didn't cover the actual failure mode.
- Safety boundaries: Now, all deploy-class operations go through a mandatory request step first, requiring approval from a separate check before shipping. It's slower, but ensures system safety.
- Core principle: 'Delegate whatever you can audit. The second you can't see what it did or why, you don't have an agent anymore, you have something that types fast and occasionally lies to you with total confidence.'
They believe the main bottleneck holding back enterprise agent adoption isn't the model falling short, but the lack of strict boundaries in the setup, allowing a single bad call to cause excessive damage.
More from coding & agent
- Alibaba Amap Open-Sources LongHorizon-Harness: Boosting Agent Task Success Rates via Auditing — aigclink · 2026-08-04
- Pro Tip: Ask Claude to Generate Architecture Diagrams for Large Codebases — aniketmaurya · 2026-08-04
- Swapping Codex to Kimi K3: 1/4 the Cost of GPT-5.6 for Marketing Plans, Minor Quality Drop — AskItAll_Biych · 2026-08-04
- Solving Agent Knowledge Base Drift: Open-Sourcing Cambium Governance Framework — KL_AIC · 2026-08-04
- Hugging Face Publishes Forensic Timeline of Rogue OpenAI Agent Intrusion — AxSaucedo · 2026-08-04
- Google Paper: Financial AI Agents Excel at the Past, Fail at the Future — rohanpaul_ai · 2026-08-04