Reddit thread: how to run long-lived AI agents without losing context or blowing up costs

Useful_Lecture_5927 · reddit · 2026-08-24

A developer asks how to manage context for long-running, multi-step tool-using agents. Convinced that keeping full history in the context window is a bad approach, they propose a combination of short-term working memory, a persistent store for key facts/results, summaries/checkpoints after certain steps, and retrieving only what's relevant to the next action — and want to know the practical sweet spot, explicitly preferring real-world experience over theory.

Original post →

More from coding & agent

coding & agent channel →