Turn Repeatable AI Agent Work Into Code
AI agents are useful for figuring out how to do something, but once the process is established, there's often no reason to keep using AI to perform every step.
Developed from a conversation between Pete Winn, Skyler Westby and Andy David
When building software with agents, Andy David takes an approach that separates experimentation from execution. An agent can help explore a problem, try different solutions and work out what needs to happen. Once that process becomes predictable, the repeatable parts can be written into conventional code rather than asking the agent to figure them out each time.
There’s a practical benefit to this. Running established processes through code is often cheaper and more predictable than repeatedly using an AI model. The agent can still manage the work from beginning to end, but it doesn’t need to handle every individual step. How much gets handed over to code depends on the task and what makes sense for the person using it.
It also changes what people need from the interface. If an agent is coordinating the work and code is handling the routine steps, there’s little reason to watch everything happening behind the scenes. What matters is being able to review the finished result, spot mistakes and make corrections. As Andy puts it, the job of interface design increasingly becomes helping people “check the agent’s homework.”