letta-code — the agent loop split across client and backend (headless, local backend) Diagram

Loop rule

each pass through 'run started' is a new backend run with one model step
the client decides to continue; the backend never loops on its own

Bounded recovery

≤ 3 overflow compactions and ≤ 3 transient retries per provider call
max_tokens_exceeded ends the turn (seen when a compacted request still fills the window)

One-shot headless

'ask' decisions are denied, not awaited
max-turns counts user turns, not approval continuations