Turn Agent Workflows Into Cheaper Software
Agents create lasting value when the processes they discover are captured as reusable software that runs reliably at a fraction of the original cost.
Developed from a conversation between Pete Winn, Paul Itoi and Andy David
Agents are useful for discovering how to complete unfamiliar work. Once they have found a reliable route, however, repeatedly asking them to rediscover it wastes tokens and preserves uncertainty. Pete Winn described the more durable role of an agent as finding what the software should be, after which the process can be encoded and run repeatably.
The economics sharpen that distinction. Pete gave the example of a process that costs $10,000 when performed by a person and $1,000 when an agent performs it with tokens. That is already a substantial saving. If another company converts the established method into software that costs $50 per run, the agent-based provider is suddenly carrying a much higher delivery cost for the same work.
This produces a clear progression from human work to agent discovery and finally to deterministic software. Humans remain available for required judgement, while agents handle the contextual decisions that cannot yet be fixed in code. Everything else becomes part of the company’s own software system, including simple interfaces that request human input only when the process reaches a genuine decision boundary.