Skip to content

Durable Orchestration

LARGESTACK now exposes run-level durable orchestration through Orchestrator(durable=True).

from largestack import Agent, Orchestrator

orch = Orchestrator(
    name="invoice-flow",
    strategy="sequential",
    agents=[extractor, validator, reporter],
    durable=True,
    thread_id="invoice-123",
)

result = await orch.run("process invoice")

Durable mode stores:

  • started checkpoint
  • completed checkpoint
  • failed checkpoint on exception
  • strategy, task, output, trace ID, cost, steps, metadata

Resume completed runs

orch = Orchestrator(
    name="invoice-flow",
    strategy="sequential",
    agents=[extractor, validator, reporter],
    durable=True,
    thread_id="invoice-123",
    resume_completed=True,
)
result = await orch.run("process invoice")

This returns the completed checkpoint without rerunning the flow.

Honest limitation

This is run-level durability, not full LangGraph-style per-node checkpoint/replay/time-travel for every graph strategy. It is a production-useful audit/resume layer, and future work can add deterministic per-node replay.