Skip to content

resil-replay

Domain: resil · Model class: cheap

Use when a user wants their orchestrator to learn from past runs and improve routing strategy over time without retraining a model. Triggers: “learn from past runs”, “workflow memory”, “consolidate experience”, “reflection loop”, “meta-learning for orchestrator”, “hippocampal replay”, “improve routing from history”, “make the orchestrator smarter over time”, orchestrator “keeps making the same mistakes”. Also trigger when someone wants a scheduled review of their workflow performance logs.

Buffer N ExecutionTrace objects (FIFO or quality-weighted eviction). At trigger, run reflection agent over buffer+current strategy. Agent outputs routing_strategy_update; inject into orchestrator system prompt.

  • “learn from past runs”
  • “workflow memory”
  • “consolidate experience”
  • “reflection loop”
  • “meta-learning for orchestrator”
  • “hippocampal replay”
  • “improve routing from history”
  • “make the orchestrator smarter over time”
  • “keeps making the same mistakes”
  • “scheduled review of workflow performance logs”
  • the user wants a one-off improvement without ongoing adaptation or structural change
  1. What traces or run logs should be replayed?
  2. How many traces and what mix of successes/failures should consolidation use?
  3. What strategy update or injection mode can modify the orchestrator safely?
  4. How often should replay consolidation run and what metrics define improvement?
  • failure mode analysis
  • recovery strategy
  • operational checks
  • validation notes

flow-orchestrator · orch-agent-orchestrator · prompt-chaining