We built Caddie in a week and learned 15 principles

A closer look at We built Caddie in a week and learned 15 principles

What this post is about

This is a short set of practical principles the Oncourse AI team used to build products with agents, written by members of Oncourse AI and titled around Caddie. It explains how agents change the work: agents now write most of the code, so the bottleneck becomes the human attention directing them.

The text frames the principles as ways to start, check, teach, and aim agents so projects do not lose coherence when many automated workers are involved.

What you can actually do by following these principles

Give tasks to agents immediately once you decide they are worth doing, so work starts instead of waiting in a backlog.

Run multiple agents concurrently (the authors recommend at least five) and assign a coordinator agent to route tasks, watch channels, and produce single summaries.

Use agents for drafting follow-ups from meetings, opening fixes from production errors, running QA with customer personas, improving work continuously with Kaizen routines, and converting successful experiments into repeatable skills.

Who will find this useful

Teams or individuals who are building software with agent-driven workflows and want a practical set of rules for organizing that work.

People who already use agents for coding, reviews, QA, or monitoring and need guidance on how to start tasks, keep quality, and capture repeatable skills.

Questions people ask about We built Caddie in a week and learned 15 principles

Does the post explain how to handle code review with agents?

Yes - it suggests making merging safe by confirming findings in the code and letting tests and the build check each line rather than manually reading every line before merge.

How should meetings change when using agents?

Meetings should start work instead of creating backlog tasks: the authors describe a 'Get it done' card that drafts follow-ups (emails, PRDs, Slack messages) and hands them to agents to be sent after your approval.

What do they recommend for improving agent behavior over time?

Turn good results into documented skills with reference examples and recipes, send agent runs through a proxy to review traces weekly, and set daily measurable goals for agents.

We built Caddie in a week and learned 15 principles was shared on NexURLs by Nexis on October 24, 2026, in Tools.