How to Ship Progressive Agents

Progressive agents, powered by large language models (LLMs), are designed to gain agency in controlled steps, enhancing their capabilities through a structured and evaluative approach rather than launching as generalized chatbots without boundaries. This method begins with a narrow, observable workflow and incrementally adds planning, tool use, memory, branching, retries, and approval gates, ensuring each capability passes realistic tests before advancing. The emphasis is on defining clear roles, inputs, outputs, and permissions, with the help of a capability matrix that outlines current statuses, required evaluations, permission levels, and rollback actions to prevent vague readiness debates. The process starts with static workflows, where tasks are clearly defined and evaluated, before introducing tools to aid in accessing or processing data. As agents evolve, evaluations map specific risks, with trace data capturing every step to identify potential failures, ensuring safe progression through stages such as static, tool-assisted, planning, and dynamic workflows. The rollout involves small, controlled cohorts with rigorous evaluations, ensuring each new capability earns its place through tests and evidence, maintaining a balance between autonomy and control.