StudioEight field lessons from building AI agents on a legacy IBM Sterling WMS
Original title: What actually took the time in an AI agent setup for a legacy ERP/WMS: knowledge, checks and approvals
The author shares months of experience building agents for a customized IBM Sterling WMS: conventions must be enforced via config rather than prompts, outputs must name evidence, designs pass JSON schema checks, scripted fact checks and an LLM reviewer, with a human (non-author) approving last. Kimi K2.5 handled triage fine, but the 22-field design step succeeded only 2 of 7 runs, so it moved to Claude Code. Undelivered knowledge caused a $3.63 design rerun that dropped to $1.75 once the fact was actually delivered; the author warns against letting agents write their own memory and moving human decisions earlier to triage and spec stages.
A first-hand retrospective from retrofitting a legacy IBM Sterling WMS, offering transferable practices on knowledge delivery, scripted checks, and moving human approvals earlier.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.