My notes on Yesterday's Claude Code Office Hours

Yesterday, Thariq from Anthropic hosted the first Claude Code Office Hours on X.

I found it very fascinating and insightful.

If you missed it, I took some notes that might interest you.

During the first part, Sid (also working on Claude Code) was invited, and he was asked by Tariq questions around his workflow.

I didn't stick around for the second part when Thariq was demoing Claude Code SDK to build an email agent.

The began with the main focus, which were the new sub-agents features.

As a reminder, sub-agents are a way for Claude Code to spawn small instances of itself, to achieve a specific task and then return to the main runner.

One of the goals is to avoid polluting the main context with unnecessary data and also allow each sub-agent to work with its own context.

This feature is mighty and has many other specific details such as scoped tool selection. Claude Code even helps you write those agents.

Anyway, Sid went over his own setup:

• One feature analyzer whose job is to understand what's going to be implemented, read code, and get to the right abstraction (Sid set up his own command to ask for the new feature to be implemented)

• One code architect / planner that takes information from the analyzer and decides how to build the feature. It output one or multiple plan, which then can be selected and tweaked by the user.

• One implementer who just writes code, Sid just accepts all edits. (YOLO mode!)

• One code reviewer.

• One code simplifier / cleaner / refactorer: his job is to try to simplify the code while avoiding breaking anything in the process.

It was interesting to hear that Sid rarely relies on Opus.

Actually, out of all those tasks, only the planner is set to use Opus. He even uses Haiku for a lot of coding tasks.

I guess those of us who are afraid of the new limits should just begin using Sonnet more.

I will try to make a sub-agents architecture for myself inspired of Sid's setup and report back my results.

They mentioned some possible improvements of sub-agents, such as a sub-agent that forks the main context (leaving the original clean), or a fire and forget one (although I didn't understand this part since I think this is already what we have).

What we have right now is just the foundation, improvements are coming, also around agent to agent communication etc.

They have many ideas, they are just careful in that they want to get whatever they ship reliable and easy to use.

Thariq and Sid talked a bit about whether using Claude Code takes anything from what makes you a good developer.

The point that was made was that it doesn't take away anything from your skills, it just allows you going faster. In the end, you're still doing the job that you always did.

For instance, the team relies on Claude Code to build… Claude Code itself!

And they're solving really hard problems while doing it.

In fact, they said Claude Code writes most of Claude Code.

When it comes to software dev, they think asking and planning with Claude Code is underrated. They gave the example of just asking Claude Code about a problem or a feature and go back and forth without even having it write the code, and it's already super useful.

• They do encourage using parallel work using git worktree (as much as your brain can handle)

• They suggest setting up notifications for parallel work, with Slack for instance (Personally I didn't have to do anything iTerm2 sends me notifications when Claude Code requires my attention and the window is not focused)

• They plan to share some command, sub-agents, and hooks examples

• They think the interruption experience is not great right now, and are thinking about making it better.

• They mentioned different use cases for Claude Code, such as going through git history of last week to summarize it for a weekly standup

• They'll have something to announce around /compact soon, and I'm super excited about that, I can't to know what it is!

I hope you found that write-up useful.