Johnny Halife, CTO of Southworks, says engineers should treat AI coding assistants like a very junior developer who knows nothing about the project or what success looks like.
Halife argues that continuous integration, continuous delivery and live deployment are now requirements because AI-built code must be verified and tested.
Southworks says its approach increased pull requests and changes tenfold on a streaming platform client's software engine, and Halife warns it must be built up gradually.
AI writes the code, so engineering judgment decides what ships
AI makes code cheap to produce, so the scarce work becomes deciding what is worth building and checking that it is right. Southworks treats AI like a junior hire. It gets clear definitions of done, curated context and automated quality gates, with humans supervising the process instead of approving every step. Without that structure, faster generation only creates a longer queue of review decisions.
Takeaways
Treat AI assistants as very junior developers. Halife says the assistant is capable but knows nothing about the project, its purpose or what success means. Engineers must define done and give tasks small enough to review yet big enough to delegate, not steer it constantly in copilot style.
Faster code generation moves the bottleneck to review. Generating code in small pieces without careful review leaves engineers mostly reviewing. Continuous integration, delivery and live deployment become requirements, since AI-built work must be verified and tried out. Cost also climbs quickly when AI builds whole apps.
Use AI to supervise automation, not sit at its centre. Halife sees AI as most useful advising on or supervising an automated process. Companies that automate checks and documentation review often see the process drift after a few months, once nobody watches the numbers.
Context is a constraint, so curate it. Everything loaded into a conversation is repeated and shrinks the usable model window, for example from 1m tokens to 50 percent after 500,000 tokens of context. Southworks uses decision records, style guides and contribution guidelines, plus MCP servers and APIs, to point agents to information.
Human in the loop means supervising through gates. Halife says the phrase should not mean a roadblock waiting for alerts. Southworks runs a pre-work checklist covering the definition of done, quality tools and cohesion of the work item, then reviews results at the end, as a manager equips a team.
Southworks built an engine that lifted output tenfold. For an unnamed streaming platform client, Southworks built an engine that maintains and builds software. Pull requests and changes rose tenfold, and documentation and quality aids improved for human engineers too. It had to be built up gradually, not requested whole.
Hiring now favors fundamentals and judgment. After the pandemic shift toward narrow specialists, Halife looks for sound fundamentals in cloud, AI and writing software, not expertise in every language. He names curiosity, communication, time management and judgment, which comes from experience and knowing what to build and why.
Tap a name to go there · tap a line to see why they are linked