Documentation index: llms.txt. This page is also available as markdown: append .md to this URL or send Accept: text/markdown.
Review Agent Work
Your agent writes Nodes as files. You decide what ships. Coalesce Desktop is where you make that call.
Why Review Matters With Agents
An agent works fast and reports with confidence. Most of what it writes is right. The mistakes that get through are the plausible ones:
- A join that fans out and doubles a measure.
- A filter that quietly drops rows.
- A dimension keyed on the wrong column.
- A rename that breaks a Node three steps downstream.
None of these fail loudly. The SQL compiles and the Node runs. Tests catch some of them. The rest take a person who knows the data, looking before the change reaches a shared Environment.
Review is cheap at this point. Everything lives as files in your Git checkout and runs in your development schema. Nothing has been planned or deployed yet.
The Review Loop
- Brief. Tell your agent what to build and how you'll judge it.
- Watch. Keep Desktop open next to your agent. Nodes appear in the graph as the agent writes their files.
- Review. Work through the four checks below.
- Fix. Ask your agent to change anything that looks wrong, then review the new version the same way.
- Ship. When you're satisfied, the agent commits, runs
coa plan, and you read the plan before it deploys. See Deploy Withcoa.
The Four Checks
| Check | The question it answers | Where you look |
|---|---|---|
| See What Changed | Which Nodes did the agent add, edit, or remove? | The Source Control sidebar |
| Check the SQL That Will Run | Does the SQL compile, and does it do what you asked? | The Node editor and its Code and Preview tabs |
| Check Lineage and Impact | What feeds this Node, and what breaks if it changes? | The graph filters and Column Lineage |
| Review Checklist for Agent Work | Are naming, grain, keys, tests, and cost right? | All of the above |
Keep the Agent Honest
A few habits make every review faster:
- Ask for a summary of every change. Have the agent list the Nodes it touched and why, then compare that list with what Desktop shows.
- Review before the agent commits. Desktop shows changes since the last commit. Once the agent commits, those changes leave the Changes list. See What Changed explains what to do instead.
- Keep tasks small. One layer or one Node at a time is easier to review than a whole pipeline at once.
- Work only in development schemas. Never let a review run touch a production schema.