Architecture AI Updates: August 24, 2026

1. A Tiered Cheap-to-Expensive Pattern for Real-Time AI Moderation

InfoQ / Bruna Pereira, DoorDash. In a QCon presentation, DoorDash engineer Bruna Pereira describes a content-agnostic moderation platform that layers three model types so cost tracks risk. Fast internal classifiers filter roughly 90 percent of obviously safe messages before any expensive LLM call, external vendor APIs handle specialized tasks such as image analysis, and LLMs are asked for severity scores across dimensions like threatening, profane, and sexual rather than a single boolean. This “cheap then expensive then action” layering lets the system apply graduated responses while holding real-time latency across millions of daily messages. Source

2. Validating Agent-Generated Changes Without Reading Every Line

Simon Willison. Writing on his weblog, Simon Willison argues that working effectively with coding agents depends on being able to confidently direct their changes and validate the results, but that validation does not require reviewing every line of generated code. He notes that eyeballing each line has never been the most effective way to confirm a change behaves correctly, and points toward outcome-oriented checks as the more reliable approach. The framing treats agent-driven development as a system-design question about verification strategy rather than line-by-line inspection. Source