What ten versions of a showreel taught me about directing with AI
I had an AI agent direct a 15-second showreel over ten versions, and it was strong wherever there was a right answer and weak wherever there wasn't.
Amadeus Lim
Software engineerKuching, Sarawak
I am a software engineer at Trade Nation. I build React and TypeScript features for a trading platform and lead the engineering team's adoption of agentic development with Claude Code. I write here about agentic workflows, context engineering and developer experience. Based in Kuching, Sarawak, open to remote work.
opens in new tabBlog
The writer should never review
Our agents kept approving their own code, so now separate agents with narrow jobs check every change in a fresh context, and a developer still has to approve it before anything ships.
We onboard agents like new joiners, with behavior docs
A coding agent needs the same briefing a new joiner gets, so every behavior decision goes into a small doc in the repo that explains how things should work, what changed and why.
Why our main agent delegates almost everything
Long agent sessions lose track of their own decisions, so our main agent plans and delegates while short-lived helpers do the reading and writing in their own clean context.
Skills, rules and hooks for coding agents
Some instructions a coding agent can skip now and then, some it has to see every time, and some a machine should enforce for it.
Now
Building agentic workflows at work, writing about what breaks, and preparing for the AWS Solutions Architect exam in December.