Operational knowledge lived in people more than repositories
Plant-specific MES configurations, ERP integrations, equipment interfaces, and years of vendor customization coexisted. Knowledge about which module controlled which equipment—and why one plant behaved differently—was concentrated among a few engineers.
Code and documentation alone could not reconstruct plant reality
Troubleshooting slowed when key people were unavailable, and morning teams spent time reconstructing overnight activity. The first goal was to use AI to map dependencies, interfaces, site-specific configuration, and test gaps in older systems.
The product set expanded as the problem changed
Patty Code and PCCP were sufficient for read-heavy analysis. Crew became necessary when findings had to become shared operational knowledge. Once specialized agents exceeded ten, Pilot was added to manage ownership, duplication, escalation, and resource use.
Workspaces and ownership were organized by system and plant
Persistent spaces were created around MES Core, Plant A, Plant B, Plant C, and the Equipment Gateway. Agents were grouped beneath the manufacturing systems lead, engineering manager, and security lead responsible for their work.
Implementation
Phase 1 · Read-only system mapping
Agents analyzed dependencies, interfaces, legacy modules, plant-specific configuration, likely unused code, and test gaps. PCCP prohibited modification of production-connected systems.
Phase 2 · Persistent collaboration records
Crew connected overnight agent analysis, open questions, engineers, and operators in one context. Shift handoffs became linked to the underlying conversations and activity.
Phase 3 · Agent organization operations
Pilot defined each agent’s responsibility, human manager, accepted work, approval boundary, escalation path, and resources. The organization could identify duplicate investigations across agents.
Observed change
- 약 9영업일
- First useful system mapFollowed by senior engineer validation
- 90–100분 → 약 40분
- Median incident triage timeCustomer observation for selected incident classes
- 30분 → 5–10분
- Morning review of overnight activityUsing persistent activity records and summaries
- 약 50%
- Regression-test preparation reductionApproximate observation in participating systems
The documented plant and the operating plant were different
An agent could produce a coherent explanation that operators knew did not match actual plant behavior. The company added a validation step in which plant engineers and operators reviewed the initial understanding before it became organizational knowledge.
Having agents was different from having an agent operating model
Pilot was unnecessary for two or three agents. Once more than ten specialized agents worked continuously across sites, an organizational layer for ownership, duplication, blockers, and escalation became essential.
Evidence and disclosure
Customer-reported and phase-observed. Results are approximate and limited to participating systems and incident classes.
Company, plant, equipment, and identifying system details are withheld. Operator statements are summarized as lessons rather than quoted.