Agents

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

Stanford / SambaNova / UC Berkeley·October 6, 2025

Qizheng Zhang

TL;DR

Treats an LLM’s context as an evolving “playbook” updated via incremental delta edits (generate → reflect → curate), avoiding context collapse so models improve from experience without weight updates.

Why it matters

Crystallized “context engineering” as a discipline and showed test-time self-improvement via context can rival weight-based adaptation.

Related terms

Context Engineering
Agentic Reinforcement Learning