بچه‌های خفن دانشگاه استنفورد تو یه تحقیق تازه نشون دادن چرا پرامپت‌های ثابت و بازنویسی کاملشون، ایجنت‌های هوش مصنوعی رو کَلّه‌پا می‌کنه و باعث افت شدید دقتشون میشه

Stanford researchers proved why static prompts break autonomous agents In a breakthrough paper on Agentic Context Engineering, Stanford researchers exposed two fatal failure modes in agent architectures: brevity bias and context collapse. Iterative full-prompt overwriting abruptly compressed 18,282 tokens of domain memory down to 122 tokens, causing accuracy to plummet from 66.7% to 57.1% Their framework replaces monolithic prompt rewrites with an evolving, delta-updated strategy playbook across three synchronized roles: 01. The Generator (Fast Execution Loop) Runs live task trajectories and tool executions in real-time. Flags which strategy bullets were helpful or harmful during inference without waiting for prompt re-optimization 02. The Reflector (Diagnostic Loop) Analyzes trajectory logs to isolate failure points. Translates raw environment errors into structured, atomic operational lessons without touching model weights 03. The Curator (Incremental Playbook Delta) Merges candidate delta bullets into the playbook using lightweight deterministic logic. Prunes redundant entries via semantic embeddings to prevent index bloat 04. Hypothesis Verification & Empirical Results Stanford demonstrated a 17.1% accuracy surge on AppWorld agent tasks and an 86.9% reduction in adaptation latency. Practical testing on production multi-turn coding workflows confirmed the hypothesis: benchmark task completion rose by 14% while token adaptation costs dropped by 70% 05. KV-Cache Serving Efficiency Longer curated playbooks did not increase serving costs. Structured prefix ordering delivered a 91.8% KV-cache hit rate, cutting billed input token expenses by 82.6% Treating context as an incrementally curated playbook bridges real-time execution with continuous self-improvement Read the complete breakdown in the article below ↓