Abstract
Agentic tools amplify what’s already in the codebase. The agent works through your code, your architecture decisions, and the documentation around them. It follows what it can read and guesses at the rest. It works well when there is one canonical way to do something, and drifts when there are five. Vibe coded greenfields hit that wall in a few weeks, and brownfield codebases got there gradually at human pace.
Variance like this has always carried a cost, but for most of software's history it was a soft one. Humans were the only contributors, they moved slowly enough to absorb it, and going back to enforce consistency rarely justified the effort. The agent changes that math, it’s the first tool fast enough to do the consistency work humans never had time for, and the first tool that actively suffers when that work has not been done. The codebase is now the agent's runtime.
On the UI Platform team at Netflix we have spent the last year retrofitting our codebase, which powers commerce experiences from signup through billing. We started by picking the canonical patterns the codebase would use. New code subscribes to them, and old code is migrated or refactored. On top of that, a context infrastructure layer runs the length of the SDLC. The team's architecture decisions seed every agent session before code is written, and the plan that comes back gets checked against them. A pre-review agent flags drift against those same decisions before human reviewers open the diff. After merge, a separate agent updates the documentation to match what shipped, so the next session starts from cleaner inputs than the last.
Each pass through the loop tightens the codebase. Patterns flow back into the code, variance drops, and reviewers see shapes they already know. When a change affects the documentation, it updates alongside the code. Platform engineering's job has not changed, but its audience now includes the agent.
Key Takeaways
- How your codebase is affecting the quality of the code agents produce.
- Why now is the time to optimize the environment the agent operates in.
- How to retrofit a mature codebase for coding agents across the SDLC.
- How architecting for agents drives codebase entropy down and code review cycle time with it.
Speaker
Mark Khuzam
Senior Software Engineer @Netflix, Previously Led the Consumer Web Platform Team @OpenTable
Mark Khuzam is a Senior Software Engineer at Netflix on the Commerce Experience Capabilities team, where he builds CLCS, Netflix's internal SDUI framework for consumer experiences across signup, billing, Games, and Ads.
His recent work focuses on agent-first architecture, the documentation layers, control surfaces, and consistency guarantees that let AI coding tools produce code matching a team's standards by default.
Before Netflix, Mark led the consumer web platform team at OpenTable. He has spent the last six years on platform teams after starting his career as a full-stack engineer.
Mark believes generative AI is as much art as it is science, and that the field still has more to discover than it has figured out.
