What we learned shipping AI agents to production
Agents are fun in a demo and humbling in production. Here are the guardrails that made ours reliable enough to trust.
A demo agent that works once is easy. An agent you can put in front of real users, every day, is a different problem entirely. This is the short version of what moved ours from “neat” to “dependable”.
1. Constrain the tools, not just the prompt
The biggest reliability wins came from narrowing what the agent can do, not from cleverer instructions. Fewer tools, stricter schemas, validated inputs.
2. Make failure observable
Every tool call is logged with its arguments and result. When something goes wrong, we can replay it — no guessing.
3. Keep a human in the loop for irreversible actions
Anything that sends, deletes, or spends gets a confirmation step. Always.
You can also embed videos anywhere inside a post:
The header video above is set with a single youtube: field in the post’s
frontmatter — that’s all it takes to turn any post into a video post.