An agentic engineering workflow for independent builders

Understand the project before changing it

Locate the entry points, storage layer, authentication boundary and build commands. Ask the agent to explain a request from the visitor’s action to the stored result. Unclear assumptions here become expensive implementation mistakes.

Ship a vertical slice

Build one complete path through the system. For a private idea board, that means create, persist, reload and enforce ownership. A dozen disconnected screens do not prove this path works.

Keep changes small enough to inspect. Use version control to preserve a working state and review the difference before moving on.

Choose predictable automation where possible

A known sequence can be implemented as a workflow; an agent is useful when selecting the next action requires judgment. Each additional agent introduces coordination and operating costs. Add concurrency only when tasks can be separated and their results checked.

For content production, separate source collection, drafting, claim checking and publication. Record the source and content revision. Retry a failed step without publishing the same revision twice. This is a workflow design recommendation, not a statement that Wantlist’s background curator is deployed.

Ask for evidence at the handoff

The final report should identify the behavior changed, the checks performed and unresolved gaps. A model saying “done” is not a substitute for the application responding correctly.

1. Inspect the code change and relevant dependency changes.
2. Run the test that would catch the original failure.
3. Try the primary journey in the actual browser.
4. Record untested external services before release.