Agentpause Suspends LLM Agents Before Rate Limits Hit

Maleficent_Pain2722 · reddit · 2026-07-20

The author built agentpause, a scheduler for long-running LLM agents that suspends work before a rate limit is hit and resumes later without losing progress.

It checks each next call against the provider’s real remaining budget using rate-limit headers plus a safety margin.
If the step would exceed the budget, it either waits just long enough for refill or checkpoints and exits cleanly.
On the next run, it resumes from the exact step instead of restarting the whole task.
For hosted APIs like OpenAI, Anthropic, and Groq, this is a logical warm start: the context is resent, but no work is redone.
On llama.cpp, the checkpoint can also save/restore the KV cache via /slots save/restore, which makes it a true warm start.
The author measured a 9k-token Qwen3-8B resume on an M1 Pro: 46.9s for cold re-prefill vs 0.5s for warm restore, about 93x faster.
An additional experiment showed that blindly truncating context can destroy recall quality, while a cheap summary preserved all planted facts.

The repo is MIT-licensed, dependency-light, and works with providers directly or via LiteLLM; it also plugs into LangGraph in two lines.

Original post →

Agentpause Suspends LLM Agents Before Rate Limits Hit

More from coding & agent