When to use an AI Coding Agent

UpdatedSeptember 21, 2026

Hi, thanks for stopping by! I am focused on designing and building innovative solutions using AI, the Oracle Database, Oracle APEX, and Oracle REST Data Services (ORDS). I hope you enjoy my blog.

Part of seriesDynamic Actions

A few weeks ago, I set up an automation in Codex to send me a weekly report on how efficiently I use AI for APEX development. The report is consistently telling me that I am using AI too much for small investigations and bug fixes, and that I should be using it more for new development and major refactors.

On this I have to disagree.

Even a small APEX bug involves the following:

Read and understand the ticket
Reproduce the issue
Isolate the code causing the issue
Plan a fix
Implement the fix
Regression test
Create a code pack
Deploy the fix to DEV, TEST, and PROD
Keep the user informed of progress

That can easily be forty-five minutes to an hour of work.

I can get Codex to do most of steps 1-6 for me and then review the diff and test results to confirm that Codex made good choices. This reduces steps 1-6 to about five to ten minutes. More importantly, Codex is going to do a better job than me of reviewing the entire codebase and finding dependencies I might miss when investigating a small ticket.

Small fixes are a big part of real APEX development, and saving time on them adds up. My experience still matters: I need to guide the agent, question its choices, and judge the result. With that oversight, even a small ticket can be a good use of an AI coding agent.

#orclapex #aicodingagents

Comments (1)

Join the discussion

Mateo Ruiz1mo ago

The distinction I’d make is between task size and task complexity. A ticket that looks like a five-line bug fix can still require understanding dependencies, deployment paths, regression risk, and surrounding code. That makes it a reasonable target for an agent even when the code change itself is tiny.

The important part is the review loop described here. The agent handles the expensive investigation and execution, while the developer remains responsible for validating the reasoning and outcome. That’s a much more useful way to measure AI coding productivity than simply asking whether the agent wrote enough new code.

Dynamic Actions

Part 1 of 7

Brief notes with my thoughts on Oracle APEX, AI, and technology in general.

Up next

Vector Search is Hard

In spite of the demos you have seen, implementing vector search for a real-world application is hard. Google Search has set user expectations for search. Those expectations carry over into APEX apps.