Tether

Durable execution for long-running LLM agents — automatic checkpoint, resume, and cross-provider failover

Tether is a Python library that brings durable execution to long-running LLM agents. Wrap your existing OpenAI or Anthropic client and get automatic checkpoint/resume, cross-provider failover, and resilient multi-step workflows — without rewriting your agent logic.

The Problem It Solves

Long-running LLM agent pipelines fail silently: a network timeout, a provider outage, or a rate limit can wipe out minutes of progress and cost. Tether intercepts every agent step, persists state, and resumes from the last successful checkpoint — making agents production-safe by default.

Key Features

  • Automatic Checkpointing — Agent state is persisted at every step, enabling transparent resume after any failure.
  • Cross-Provider Failover — If OpenAI fails, Tether automatically retries the same call against Anthropic (or any configured fallback) without dropping context.
  • Drop-In Client Wrapper — Wrap openai.OpenAI() or anthropic.Anthropic() with a single line. No agent rewrite required.
  • Idempotent Replay — Replayed steps skip already-completed work, saving tokens and cost on recovery.
  • MIT Licensed — Production-ready, open-source, self-hostable.

Usage

from tether import TetheredClient

client = TetheredClient(primary="openai", fallback="anthropic")
# All agent calls now have automatic checkpoint/resume + failover
response = client.chat("Analyze this dataset and return statistical findings")

Tech Stack

Python · OpenAI API · Anthropic API · SQLite (checkpoint persistence) · MIT License

GitHub