Onboarding SubAgents is killing us.

Chris,

What newsletter about tech would be complete without a diatribe about AI and how it's reshaping the industry. I'm going to look at it through the lens of what could be an entire blog itself. It's doing a lot of shining lights on bad practices people have had and making them obvious.

"Lets begin at the beginning

we're lovers and we're losers

we're heroes and we're pioneers

and we're beggars and we're choosers

we're skirting round the edges of the ideal demographic"

Frank Turner (I knew Prufrock before he got famous)

Onboarding is crashing into individual developers like the meteor that killed the dinosaurs. Setting up Agents and wondering why the hell they did that I find the most hilarious thing. You've not onboarded them, you've given them a rough outline of what is to be done without giving them a why. The infuriation at when they come back with their own take on what you wanted makes me remember every onboarding experience I ever had.

Remembering that each agent that you spin up is a newbie onboarded is something that is being overlooked pretty wholeheartedly. We give them just enough information to get the job done without any sort of context. This saves on tokens but the product is easily suffering.

A well rounded Mission Statement, a why the company exists overall, would give a great amount of context to the agents. A Vision statement or where the company is trying to head what it's focused on and what it isn't is the next step there. The biggest thing a company can do to help its employees is tell them what they don't need to work on so that there are fewer fields for them to wander into.

Creating a style sheet is something that can easily be done without the assistance of AI saving on tokens and cost as well. This will also save time on arguing with your claude instance over the merits of tabs versus spaces. Though you can still do that if you're bored at 3 am with insomnia.

The lesser known things that help out creating new employees are cultural problems which I'm hearing reports is just now getting into the limelight. "We always run tests cause otherwise we feel like failures" is a prompt I've heard to keep the AI from skipping testing altogether. This doesn't need to be so dire for a human but it still needs to be said and shown.

Questions for you on this stub of an article Chris. Where do your cultural rules live? As AI sub agents become more and more of a thing how are you making explicit the implicit things it should know?

Apologies for this being so short and stubbed but it's being written on a Monday at 10:15 and something is better than nothing.

I hope this helps

Mark.