A scheduler that pauses agents before rate limits hit
Maleficent_Pain2722 · reddit · 2026-07-20
A Reddit post introduces agentpause, an MIT-licensed scheduler for long-running LLM agents.
What it does:
- Checks the estimated next-step cost against remaining rate-limit budget before every LLM call.
- If the call would exceed the budget, it waits only as long as needed or checkpoints and exits cleanly.
- On the next run, the agent resumes from the exact step instead of redoing work.
The author also distinguishes logical resume from a true runtime warm start:
- Cloud APIs can only resend the full context or use provider-side prompt caching.
- With llama.cpp, checkpoints can include KV-cache via /slots save/restore, which skips re-prefill entirely.
- On an M1 Pro, resuming a 9k-token Qwen3-8B context dropped from 46.9s to 0.5s, about 93x faster.
Other findings:
- History compression can be counterproductive because it breaks prefix caching.
- Blind truncation can destroy factual recall and sometimes causes the model to hallucinate replacements.
- The repo includes benchmarks, works with multiple providers, and plugs into LangGraph with minimal setup.
Related event: agentpause: Scheduler Pauses Agents Before Rate Limits(2 posts)→
More from coding & agent
- Team demos AI agent Blender skills with impressive 3D results — noahsolomon · 2026-10-26
- A 13KB Markdown file can hijack Claude Code and Codex via your Downloads folder — pbaylies · 2026-10-26
- A Fallback for an Auto-Posting Bot: Silence Is the Default, Posting Is the Exception — petrusenko_max · 2026-10-26
- Claude Code can now edit videos: a 10-prompt workflow from raw footage to final cut — TawohAwa · 2026-10-26
- Developer Moves All Prompts to Skills: AI Coding Workflow v3 Explained — rseroter · 2026-10-26
- charlie Marsh: rewrites beat external deps in the AI era — mohamedmansour · 2026-10-26