90/30 Club (ML reading) #27: Agentic Context Engineering

​Week 27: Agentic Context Engineering

​Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

​This session explores Agentic Context Engineering, a new paradigm for enabling language models to modify and evolve their own prompting environments. By treating context as a manipulable state rather than static input, models can iteratively refine their behavior, reasoning scaffolds, and goal representations across tasks. We’ll examine how contextual memory, reflection loops, and tool use contribute to self-improvement, bridging the gap between prompt tuning and autonomous agency. Discussion will focus on mechanisms for maintaining alignment and stability as contexts become increasingly agentic.

​Join us at Mox to explore:

- What are "brevity bias" and "context collapse" and why do both represent fundamental problems?

- What fundamental assumption about LLMs does ACE challenge, and what evidence supports it?

- Why does ACE use three separate roles (Generator, Reflector, Curator) instead of one model doing everything?

​🔎Analyzed Papers

​​Discussion at 20:00, (optional) quiet reading from 19:00.