We've had software engineering principles for decades. SOLID. YAGNI. Lean. TDD. All earned through pain.
LLM agents are different. The principles are still forming. I want to catch them early.
"Success Is Silent" is one to remember.
Any tool or check that passes should produce no output beyond its exit code. Nothing. Fail? Dump everything.
Unix already knew this. cp says nothing on success. grep stays quiet with no matches. For agents it matters more. Context is finite.
Every ✓ compiled and 0 errors is a token that isn't there when something actually needs it. Enough of that and the agent hits the dumb zone too early.
Dexter Horthy's article has a practical solution: a run_silent wrapper. Wrap a command in it: success gets a single ✓, failure gets everything.
Worth stressing: this belongs in the harness. Hooks, wrappers, CLI output design. By the time a prompt instruction tells the agent to ignore output, those tokens are already spent.