Tea. Earl Grey. Hot.

“Tea. Earl Grey. Hot.”

We know why Agentic AIs fail in large, legacy repos. Ask a principal engineer or a frontier AI the same riddle: “I have three sons. Two of them are named Sean and James. What is my third son’s name?” If the answer isn’t in the text, all you can do is guess.

The same is true of legacy repositories. A principal engineer or frontier AI can read every line of code, but the product requirements, the business rules, the tradeoffs — are those in the comments? Are you asking the AI to guess?

With Agentic AI, our role as software engineers has the opportunity to change. Think of a new feature as a cup of tea. We have spent lifetimes learning how to prepare the leaves, what makes a tasty cup. All the tradeoffs and questions to ask product and our customers’ tastes — we spend years learning what makes a good cup of tea at our company, and then our job as software engineers is to make a perfect cup.

Our opportunity is to stop making cups of tea by hand, and turn our dev environments into replicators: feature specs go in, verified features come out.

To do this in anything outside a toy repo, you can’t rely on the agent to guess right, you need to encode these things. That means your CLAUDE.md’s (or AGENTS.md’s if you’re of that persuasion) need to say more than just what the test binaries are. They need to explain what your company does. They need to include information like what your custom jargon all means, so when you say ‘customer’ it knows ‘not the users on the free tier’. Then all the stuff only we know — don’t let your agent guess what makes a good test. Definitely don’t let your agent guess what “verified” means!

My CLAUDE.md files all say: “The feature’s not verified until you launch Claude in Chrome, and prove the feature works in the dev environment’s UI, making new test data if necessary.” Yours should too!

As software engineers, when product asks for a new feature, we get out the kettle. Instead, we must start by stuffing the product requirements into our agents, and start the long work of turning dev environments into replicators.

Then our craft evolves from steaming the water, mixing leaves, breaking out the kiln and turning clay into a cup — to making a system that takes product specs and pops out verified, working features. Like it’s the year 2360 and we’re in Star Trek. Build your replicator, tune it to your company, and then just order what you want: “Tea. Earl Grey. Hot.”