The most underrated artifact in AI engineering is a folder of markdown files.
Before any agent touches code on a brownfield engagement, the team builds a knowledge graph: every module, dependency, data flow, and domain term documented.
The agent reads this context before writing a single line. Every time.
On one of our logistics clients, building the knowledge graph took 4 days. It saved roughly 4 months.
The agent stopped generating code that violated naming conventions. Stopped importing deprecated modules. Stopped wiring services through patterns the team abandoned six months earlier.
At another healthcare client, the same knowledge graph served double duty. AI agents used it for code generation context. Human engineers used it as onboarding documentation.
I mean, most teams skip this because it feels like overhead.
Four days of documentation before any code ships feels slow.
But the teams that skip it spend those same four days spread across four months of rework, reviewing PRs that look right and break something upstream.
We've seen this at
@LimestoneHQ on dozens of engagements.
The knowledge graph is the scaffolding everything else stands on.