the concurrency data is interesting. been running parallel agents with git worktrees and the bottleneck is usually context window limits before actual compute. curious if you saw similar patterns in the telemetry
mostly i use CLAUDE.md files as persistent memory between sessions. so even after /clear or starting fresh, the agent picks up where it left off because the project instructions carry forward. for longer tasks i also break things into smaller scoped sessions instead of trying to keep one giant context alive. not perfect but it keeps things moving without losing too much
the concurrency data is interesting. been running parallel agents with git worktrees and the bottleneck is usually context window limits before actual compute. curious if you saw similar patterns in the telemetry
Hi @pvd, we haven’t studied token budgeting as a limitation in this piece. But something we will be looking in future research.
When you hit the context window you do /clear or something else to keep working ?
mostly i use CLAUDE.md files as persistent memory between sessions. so even after /clear or starting fresh, the agent picks up where it left off because the project instructions carry forward. for longer tasks i also break things into smaller scoped sessions instead of trying to keep one giant context alive. not perfect but it keeps things moving without losing too much