Inside Uber Eats' Self-Tuning Multi-Agent System for Photo Processing
prashantkr_00 · reddit · 2026-08-03
To automatically fix poor merchant photos across 10,000+ cities without generating fake "AI slop," Uber Eats built a multi-layered multi-agent architecture.
Core Architecture & Collaboration
- Uses a "Swiss Cheese" model: a Router Agent decides on processing, Editor and QA Agents loop to refine, and a Publish-Ready Gate does final safety checks.
- The QA Agent detects hallucinations via pairwise comparisons and feeds explicit instructions back to the Editor for retries, capped by a Pass@K metric.
Self-Tuning & Logging
- All agents log micro-decisions into a single flat JSON trace for easy debugging.
- A Diagnoser Agent analyzes production logs and human feedback to automatically rewrite prompt configurations, auto-deploying after passing a "Golden Dataset" benchmark.
More from coding & agent
- Non-Developer Shares Project Built Over Months Using Claude — X-ego · 2026-08-03
- Dev Modifies llama.cpp Source to Fit Agent Memory System — Savantskie1 · 2026-08-03
- 5 Engineering Lessons for Building Reliable Local AI Wake Words — InternationalGap3698 · 2026-08-03
- Parlor v2 Open Source: Local GPT-Live Clone on M3 Pro — ffinzy · 2026-08-03
- Tools Encode Trust: Stack Overflow on Dev Tooling in the Agentic Era — jedisct1 · 2026-08-03
- GraphArc: Open-Sourcing Real-Time Visual Orchestration Graphs for AI Agents — Desperate-Ad-9679 · 2026-08-03