Current AI agents operate on-the-fly, synthesizing plans and executing actions in seconds. This paradigm bypasses disciplined software engineering processes like testing, adversarial evaluation, and staged deployment, leading to brittle and vulnerable systems that can fail in high-stakes scenarios.
The authors propose an AI Workflow Store: a repository of hardened, deterministically constrained workflows produced through automated software engineering and reused across users. By selecting vetted workflows rather than improvising per prompt, agents can achieve far greater reliability and security, at the cost of extra compute that is amortized through reuse.