The Agent Architect #31: Sliding Window Management

今週のパターン

Sliding Window Management

概要:

実行中にコンテキストウィンドウを動的に縮小・拡大し、最近かつ関連性の高い情報を優先しつつ、重要度の低い古い内容を切り捨てる。

使いどころ:

会話履歴の全体がトークン上限を超える長時間のタスクで、周辺的な詳細を落としつつ、タスクに不可欠な情報は保持する必要がある場合。

注意点:

関連度のスコアリングはヒューリスティックに基づくため、後で必要になるコンテキストを誤って捨ててしまい、エージェントが同じ作業を繰り返したり、重要な制約を失ったりすることがある。

今週のエージェントAI

• A corrective agentic hybrid RAG and an operations-grounded evaluation for a scientific facilityarXiv cs.AI

  APS-RAG is a deployed retrieval-augmented generation platform that integrates decades of operational knowledge from scientific facilities including logbooks, documents, and control-system data.

• Agentic Permissions Policy Algebra for Taint Confinement in LLM AgentsarXiv cs.AI

  APPA is a permissions policy algebra for LLM agents that enables selective taint tracking to balance security against prompt injection while maintaining utility.

• Opaque Epistemic Mediation: How LLM Deployment Configurations Shape the Validation of Pseudo-SciencearXiv cs.AI

  Four major LLM families show inconsistent and non-transparent stances on contested scientific claims across multiple time snapshots.

• CausalForge: A Formally Grounded, Self-Improving Agentic Framework for Automated Research in Causal InferencearXiv cs.AI

  Framework automates causal inference research by using formal verification instead of LLM reviewers to evaluate candidate results.

• TRACE-ROUTER: Task-Consistent and Adaptive Online Routing for Agentic AIarXiv cs.AI

  Router selects LLMs for agentic workflows by considering task-level outcomes rather than making independent per-call decisions.