The Steelman: When an AI Agent Actually Earns Its Complexity

In a follow-up to his critique of 'AI agents' as glorified deterministic pipelines, James Anderson explores the narrow conditions under which autonomous AI systems are truly justified. Anderson argues that autonomy is a significant cost, introducing nondeterminism and debugging challenges. He proposes a 'steelman' framework requiring four conditions for genuine agency: the environment must be unpredictable, the path must be discovered rather than designed, the autonomous surface must be minimized to specific decision points, and outcomes must be easily verifiable with an audit trail. Most systems currently labeled as agents fail these criteria, functioning instead as rigid pipelines. Anderson concludes that while true autonomy is powerful, it is rarely necessary. He challenges developers to identify production-grade systems that meet his rigorous standards, emphasizing that unless a task is genuinely dynamic and verifiable, developers should prefer deterministic code over the complexity of an agent.