Why orchestrator LLMs also go OOD: how CodeAct/MemGPT/RLM curb context rot

a1zhang · x · 2026-11-20

a1zhang explains the paper's motivation: most tasks are decomposable (like MapReduce), but mainstream harnesses use "trajectory-as-a-prompt," stuffing all history into context — and LLMs generalize unreliably, causing context rot. Crucially, even the LM orchestrating sub-agents can go OOD, a common failure of long-running coding sessions. Harness abstractions like CodeAct, MemGPT, RLMs, and CLMs avoid this by enforcing how context moves and what each LLM call sees.

Related event: MIT Paper: LM Harnesses Are Compositional Generalizers, Not Crutches(10 posts)→

Original post →

More from coding & agent

coding & agent channel →