Agent OS: managing AI agent instructions as a versioned system, not loose prompts

A common problem with internal AI assistants is that their instruction files grow uncontrollably over months, accumulating contradictory rules with no clear precedence or way to test for conflicts. The proposed solution, called Agent OS, treats the full set of agent instructions as a versioned, schema-driven system rather than a single monolithic text file. It separates instructions into layers — a small always-loaded core, task-specific workflows loaded on demand, and reference materials triggered only when a workflow requires them — preventing unnecessary context inflation. A declared precedence order ensures hard restrictions always override user preferences or session memory, eliminating inconsistent behavior across runs. Routing between workflows is handled via an explicit decision tree with testable triggers, rather than natural-language instructions, allowing the team to validate behavior against a benchmark of positive, negative, and ambiguous cases.