i build many apps. i also work at ibm building ai. i have experimented with many different ways of working. and im trying to identify a sweet spot. - level 1: manual coding. lol no one does this anymore - level 2: agentic coding: using claude code/codex/whatever, but in a single-threaded sequential conversation-i describe what i want, it cooks, i review, we ship. an extension of this is spec coding where i describe specs or PRDs and it does it after grilling me and i review and we ship - level 3: parallel agents: i start multiple claude code sessions and say "in a separate branch and separate worktree, build this". there are like 7-8 sessions in parallel. they each finish independently, i review them independently, and then we merge it all into the main branch and worktree and deploy at the end - level 4: full software factory: i create a linear workspace and work with an agent to come up with a full roadmap-projects, epics, tickets, etc. and we fill it. then, i create many agents. one called "pm", one called "maintainer", and as many engineers as i need, and coordinate them. claude code is cool because it lets one agent talk to another via @ -mention, so i just tell pm to tell @ -maintainer to do something etc. i then talk exclusively to pm and its all coordinated. for daiku i did exactly this and told pm to "finish the entire linear board" as a goal and it has been working for 3 days nonstop and will keep going. i am paying a fortune in tokens but its worth it because its for science lol. this is working quite well, better than i expected - level 5: ??? whats next? u tell me from this ive noticed some tradeoffs: - level 1 is obslete - level 2 is tool slow - level 3 is kinda nice - level 4 is very hard to keep track of in my human context window and i often find myself happy because its moving forward, but also kinda confused (i have impostor syndrome and fomo in my own project) because i dont really know where it is and whats happening and whats the last few things that happened. i often have to ask for an update, which is so weird. also when things break, i just tell them to fix it and they did but i dont really know more after all this, i would probably choose some hybrid between level 2 or 3 depending on roadmap and timeline. i find the "separate worktree separate branch" thing kinda annyoing to setup, teardown, and clean up, so i tend to bias towards level 2 even though its soemhwat slower. maybe also because it simulates actual human communication and the feeling of community/pair programming whihc is maybe what we all really want at the end of the day anyway wdyt?