Agent execution rules
Giving an agent tools is the easy half. The rules for using them are the half that decides whether the output is worth reading.
Two of them carry most of the weight here:
Consult before deciding. Before choosing a pattern, picking a technology,
designing an API or implementing a non-obvious algorithm, query the corpus with
rag_context and a description of the task, or with the narrower rag_*
tools when you know what you are after. The corpus
exists so that architectural decisions are made against written-down experience
instead of a plausible-sounding guess. Weak results are information too: they
mean the corpus does not cover this, which is what mx rag gaps collects and the
research pipeline acts on.
Evidence before assertion. No claiming a fix without running it, no “should work”, no marking a task complete without the command output that proves it. This is the same standard the testing page holds CI to, and the reason defects live in public in the known-broken lane rather than getting quietly written around.
The full rule set (RAG tool routing, functional design foundations, error handling strategy, the procedure agents follow) is a corpus document, so agents retrieve it the same way they retrieve everything else:
It is long, and deliberately so: it doubles as the index into the theory and pattern material the rest of the corpus holds.