I wanted an agent to sort my email. When a message was a configuration question or a technical fault, something had to answer it. That something needed to know how each customer is set up, what the platform can do, and where its own knowledge ends.
System Model
A model of how a software platform is configured for each customer, built from exports and documentation, served read-only to AI agents. It answers configuration questions with a source, and says plainly when it does not know.
- Python
- MCP
- deterministic build
- read-only serving
The story
-
Where it started
-
The unfinished piece
Months earlier we had started skills that turn a customer's live configuration into a readable document in a fixed template. Feeding them the company documentation worked well for config-to-document, and only partly for explaining how the platform behaves. The document work and the need for a technical brain became one model.
-
What broke first
The generator worked on day one. It also exposed the real mess: settings with no written home, two correct sources that sounded like a contradiction because they spoke at different altitudes, and one word that meant three things depending on who said it.
-
The failure that taught the most
A real question came in and the model hedged, while the right answer was already in the repo. It had searched what customers already have, but the fix was a capability that nobody had switched on yet. The search was looking in the wrong place. Catalogue-first search came from that day.
-
Kitchen and plate
Raw exports and documents stay in the kitchen. Deterministic scripts cook the same dish every time. Agents only get the plate: an allowlisted, secret-scanned bundle served read-only over MCP. If it is not on the list, it is not served.
-
What it can do
- Resolve a setting through the override layers and say which level won.
- Search capabilities no customer has used.
- Draft a standard overview with unknowns left as visible blanks.
- Route a missing answer to the right person and fold the reply back with who and when.
The rules that came out of it
- Zero guess: every fact carries a source and a confidence level. Unknowns go to a ledger.
- A customer's own export beats a documented default for that customer's actual value.
- For meaning, the newest sourced document wins; the old rulebook is a floor, not an authority.
- Rare is not wrong: a deviation is a place to look first, never a defect.
- Unprecedented is not unavailable: a capability no customer uses is still a valid answer.
- Search before you file a gap.
Honest limits
It is a snapshot, so it cannot say what changed last week. Behaviour coverage is thin outside a few flows. Diagnosis is not solved. Mostly I have been its user so far.
What it is not
Not a continuous "company brain" that swallows chat and hopes. Not a repo-history compiler. The hard part is upstream: authority, contradictions, provenance, and the interview loop that fills what exports never contain.