After Building 400+ Production AI Agents, This Engineer's 12-Layer Stack Explains Why They Fail
MaryamMiradi · x · 2026-09-17
Maryam Miradi distills lessons from building 400+ production AI agents: tools come and go, but failures trace back to the same missing engineering layers. She stress-tested the idea by deliberately toggling layers on and off in a Medical Emergency Department triage agent system.
Her 12-layer stack starts with:
- Business Understanding: define decisions, process, success criteria, and human roles — e.g., exactly what the triage agent may decide
- Data Understanding: measure completeness, availability, quality, and inconsistencies (was the allergy missing or never loaded?)
- Knowledge Engineering: ingest, structure, index, connect, and retrieve domain knowledge
The takeaway: AI coding makes building agents trivially easy, but durable systems need the full 12-layer engineering context — not new tools or architectures.
More from coding & agent
- Open coding harnesses' underrated edge: mixing multiple frontier models — mitsuhiko · 2026-09-17
- Short film made entirely by AI agent 'astra' controlling Blender for VFX — OpenAIDevs · 2026-09-17
- PiSSA-style SVD init for LoRA adapters also improves downstream RL training, Trajectory Labs reports — simonguozirui · 2026-09-17
- Feeling Micromanaged by AI Agents? Your Job Is Defining What Good Looks Like — alex_frantic · 2026-09-17
- Copilot CLI v1.0.86 Lets Custom Agents Ingest AGENTS.md and Other Repo Instructions — copilot-cli-release-app[bot] · 2026-09-17
- NetEase Youdao open-sources Confucius4-R2T2, append-only streaming ASR built on Qwen3 — rohanpaul_ai · 2026-09-17