XCON was an expert system that configured Digital Equipment Corporation's computer orders — matching components, checking compatibility, generating floor plans. It went into production in 1980 with about 700 rules. By 1987 it had grown to 6,200 rules drawing on roughly 20,000 parts, and it worked: DEC's engineers reported "no problem with XCON's performance."
Keeping it current was the problem. Fifty percent of those rules changed every year — around 3,100 modifications annually. Each change required understanding how one rule interacted with thousands of others, and the people who'd written the originals had often moved on. Rule rationale lived in someone's memory, or nowhere. The engineers noted that update costs "do not grow linearly." The codebase became what they called "a rat's nest of special rules, tightly coupled rules," while still producing correct output.
Agent systems that maintain representations of APIs, form fields, and business rules across target platforms face the same grind. The external world doesn't hold still while you figure out what broke.
XCON by 1987
- ~6,200 configuration rules (up from ~700 in 1980)
- ~20,000 parts in its database
- In continuous daily production for seven years
The churn math (50% annual turnover)
- ~3,100 rules modified per year
- ~60 per week
- ~12 per working day
What went wrong
- Rule interactions became unpredictable at scale
- Original authors moved on; rationale was undocumented
- New developers copied existing rules and edited them without understanding every function
- Engineers grew "afraid to modify the rules, lest some unwanted behavior might result"
DEC's fix
- Reorganized the system into 40 explicit problem spaces, making implicit relationships visible
- Open question at the time: "what happens when the rule base hits 18,000 rules? 27,000 rules?"