Hey everyone,

I have been running setup with an AI coding subscription and then tried scaling into many parallel autonomous sessions. Even after all the token saving tricks, I still hit the wall hard.

I switched to lower effort modes, aggressively compact sessions around 200 / 300k, limited MCPs and skills, added a knowledge base, and built these autonomous loops (brainstorm → spec → plan → subagent work → recollection). Each one runs in its own worktree and I manage everything through tmux on a remote machine.

The funny part is I reduced tokens per session hour quite a bit, but because I started having way more sessions for longer, my total daily usage went up like 6x. I optimized the process but also I created way more work than one subscription could handle.

Also now the biggest headache is merging. When you have multiple sessions running in parallel, getting everything back to main without conflicts or drift becomes painful. I want to keep real momentum across many things at once without constantly babysitting merges or waiting on limits.

I'm currently messing with a passive "judge" system that collects simple reports from each session during compaction so I have better cross session awareness (repo here: https://github.com/muslewski/agentic-sage). Also looking at hybrid model routing and maybe running some local models for parts of the work.

For people running many autonomous sessions - what actually helped you once one high-tier plan + optimization wasn't enough anymore? Routing setups, hybrid strategies, local models that are good enough, or any tools for managing multiple worktrees?

Would appreciate any sugestions. Happy to share more details on my current setup too.

Thanks!