i have 4+ terminals open right now. each one is running claude code (in bypass permissions). each one is working on a different project.

2 of them are running experiments in independent worktrees, and the other 2 which im working on closely to debug and fix a few prod issues.

i’m not doing any of this. i’m just… orchestrating.

the old advice #

there’s a popular take that calls context switching “the silent killer of productivity.” the advice: batch your tasks, use do-not-disturb, single-task your way to clarity.

and for years, this was right. context switching was expensive because you were the one holding all the context. deep in a problem, someone pings you, and you’ve lost 23 minutes rebuilding your mental model.

but that advice assumes you’re the one doing the work.

what if you’re not?

the shift #

harness writes code now. claude code, cursor, aider, open code - the fs they operate, hold context better than you ever could.

so your job changes. you’re no longer the executor. you’re the execption handler - just looking at the logs and steering the way you want it to work.

think about it: the expensive part of context switching was rebuilding your mental model of the codebase, the bug, the feature. but if harness is doing the work, it holds that context. your job is just to feed it enough information to start.

and gathering context? that’s cheap to context-switch. “here’s the repo, here’s the bug report, fix it” takes 30 seconds. you don’t need deep focus to write a good prompt.

the workflow:

1. terminal 1 → claude code → “experiment X.. "
2. terminal 2 → claude code → “write tests for the …”
3. terminal 3 → claude code → “debug this production error, use this mcp to fetch logs…”
4. terminal 4 → claude code → “generate a html for this service internals…”

each agent holds its own deep context. you rotate between them - checking progress, answering questions, providing more context when needed.

the context switching still happens. but it happens in your terminals, not in your head.

how i actually do it #

i have a small notes i maintain in my apple notes app, nothing complex just a list of things that needs to be done by end of the day.

i now just drop each of this experiment i want to run in individual cc session, which i always start wiht plan mode, which is my way of dumping all the context and thoughts before leaving the harness do its thing. once the plan is confident enough to get me some learnings + outcomes, ill look into other sessions which might need more of my attention into steering it in better directions to find the right issue.

a few skills that help me in doing this all day:

• gstack
• manual repo level skills ive created
• a few global skills + my way of executing things (my mental model of approaching things)

im not opposing that context switching is bad, but the limit of humans context switching can be pushed when working with these agents.

thanks!