AI Architecture Updates: September 19, 2026

1. DoorDash Put a Price Tag on Agentic Tech-Debt Cleanup: $4.79 per Flag

DoorDash built a multi-agent LLM system to retire stale feature flags across more than 60,000 flags and 623 repositories, and published numbers rather than a narrative. In an evaluation of 50 flags, 45 produced usable pull requests at an average of 13.8 minutes and $4.79 per cleanup. The architecture is the interesting part: live experimentation data reaches the agents through MCP, each agent works in its own isolated Git worktree so parallel runs cannot collide, an engineer approves before anything merges, and validation is automated. It is a concrete answer to what agent fan-out looks like when the task is mechanical and the blast radius is a repository. Source

2. A Four-Layer Pattern for Facial Verification at Three Thousand Concurrent Users

InfoQ published an architecture walkthrough for high-volume face verification, starting from the failure mode that synchronous API calls collapse when three thousand employees verify at once. The design has four layers: client-side filtering that cut cloud costs 30% by discarding unusable frames before they leave the device, decoupled detection and verification that enabled 10x scaling by letting the two stages scale independently, risk-based dynamic thresholds instead of one fixed confidence bar, and zero-trust privacy with consent gates and automated data purging for GDPR and HIPAA. Source

3. Matt Pocock on Why Engineering Fundamentals Got More Load-Bearing, Not Less

The Pragmatic Engineer published a conversation with Matt Pocock on how he uses AI coding skills and agents to plan and build software. His position is that the fundamentals matter more than ever under agentic workflows, which is the opposite of the usual framing that agents let you skip them. Source