How senior AI engineers architect production agents around failure detection
MaryamMiradi · x · 2026-09-09
Maryam Miradi outlines her layered architecture for production AI agents that are measurable, traceable, and recoverable — built on the assumption that agents will fail, and the system must detect, explain, recover from, and prevent failures.
- Foundation stack: pick one primary agent framework before adding evals, tracing, or recovery; define typed state, tool schemas, agent boundaries, and MCP interfaces up front; keep the model replaceable. Suggests Claude Code for implementation, LangGraph/Pydonic AI/Google ADK for runtime, MCP for standardized tool access.
- Evaluation: define success before tuning prompts or models; test final answers and agentic behavior.
- Later layers (truncated in the post) cover failure detection and recovery — the guiding principle is designing for inevitable failures.
More from coding & agent
- MacStories' Viticci open-sources Apple Frames CLI for agent-driven screenshot framing — rudrank · 2026-09-09
- Open-source 1.6T-param MoE Nex-N2.5-Max lands 0.1 points behind Claude Opus 5 on agent benchmark — airesearch12 · 2026-09-09
- Muse agent is deliberately decoupled from credentials and network permissions for safety — rohanpaul_ai · 2026-09-09
- Box launches Mount to sync Box folders into agent sandboxes with built-in governance — badphilosopher · 2026-09-09
- Dev wires bug reports into an MCP server so his AI assistant can triage them itself — Weak_Needleworker_47 · 2026-09-09
- Anthropic interviews WisprFlow, Actively and Pendo on building with Claude Managed Agents — ClaudeDevs · 2026-09-09