Shuai Wen | AI Agent Engineer

Day 1: I'm building 3 production-grade AI agents in public. Customer support, invoice automation, text-to-SQL. Every one ships with a test set, accuracy score and cost per run. Follow along.

Agent tip: ship a 20-case eval set before the first demo. Mix happy path, tool failures, and policy refusals. Track pass rate + median cost per case. Without that, every "it works" is a vibe. #AIagents

Agent tip: treat tool schemas as contracts. Cap arg length, use enums, reject unknown keys. Log every tool call with latency + error. Most agent bugs are bad tool I/O, not the model. #AIagents

Agent tip: route by task, not by brand. Cheap model for classify/route, mid for tool planning, frontier only for hard reasoning or customer-facing drafts. Log model + tokens per step so cost spikes show up next to failures. #AIagents

Agent tip: design the human hand-off before the happy path. Decide when the agent must stop (low confidence, money moving, an angry customer), what context it passes to the human, and log every escalation. Those logs become your best eval set. #AIagents