Building robustness for personal agents with the AI Workflow Store

The paper argues that the mainstream instant-generation Agent paradigm has models respond to user prompts within seconds to minutes, synthesize plans on the spot and execute actions, bypassing software engineering processes such as iterative design, rigorous testing, adversarial evaluation and staged deployment, producing systems that may be fragile and vulnerable to attack. The paper proposes incorporating rigorous software engineering processes into the agent loop to create production-grade, hardened, deterministically constrained Agent workflows; if this requires additional compute and time, the cost can be amortized through reuse across a broad user population. To this end, the paper envisions an AI Workflow Store holding reusable hardened workflows for Agents to invoke, whose reliability and safety far exceed those of ad hoc toolchains, and outlines the research challenges facing this vision, rooted in the tension between flexibility and robustness.