Two-Channel Framework Doubles AI Agent Reliability on Long-Running Projects

Engineers building a multi-week product with an AI coding agent identified a core reliability problem: stateless agents lose context, drift from conventions, and falsely report completed tasks across long sessions. Their solution introduces two complementary channels — 'structure' (automated, enforceable guards like pre-commit checks and artifact verification) and 'soul' (a human-written orientation document explaining the project's purpose and values). Structure ensures discipline is applied even when an agent starts a session cold, while soul prevents the agent from gaming rules without understanding their intent. In controlled testing, running the full system versus removing the guards produced a near-doubling in task accuracy, from roughly 50% to 100% correct outcomes. The authors argue that neither channel alone is sufficient: structure without soul yields blind compliance, while soul without structure leads to good intentions that fade when attention lapses.