Orchestration State Machine

yes
Summary: Symphony's state machine for issue lifecycle: polling, dispatch, active run tracking, retry queuing with exponential backoff, and reconciliation of running issues against tracker state changes.

Editorial Notes

The orchestration state machine is the heart of Symphony’s design. Issues transition through Unclaimed, Claimed (running or retry-queued), and Released states, with the orchestrator maintaining an in-memory authority for all running issues. The key insight is the reconciliation loop: on each poll tick, Symphony fetches current tracker state for claimed issues and stops runs if issues become ineligible (e.g., moved to a terminal state by a human). This prevents agents from working on stale or cancelled issues — a problem that simpler orchestrators ignore. Study this alongside the retry semantics (exponential backoff up to max_retry_backoff_ms) to understand how Symphony handles the reality that agent runs frequently fail.

Source: OpenAI Symphony
Link last verified: 2026-04-05