#LLM Agent
7 related articles
TrawMem: Reframing Agent Memory as Task-Oriented, Role-Adaptive Workspaces
TrawMem proposes a task-oriented, role-adaptive workspace framework that reframes agent memory from relevance-based retrieval to organization by task roles and relations. Macro-F1 improves 2.23–3.14 points on LoCoMo and MemGallery while reducing context tokens.
Building a Search and Fetch Layer for LLM Agents: Caching, Content Extraction, and Cost Control
Learn how to build a search and fetch layer for LLM agents: use a gateway pattern to unify caching, budget control, and robots.txt compliance, extract content with trafilatura to compress 180KB HTML to 6KB, and achieve a measured 35%-45% cache hit rate to significantly reduce agent fetch costs.
LLM Agent Task Decomposition: The Core Engineering Pattern for Building Reliable Workflows
Structured task decomposition is a key engineering pattern for reliable LLM Agent workflows. Learn how LLM-driven prompt design breaks complex goals into atomic sub-tasks, reducing hallucination, improving debuggability, optimizing costs, and enabling parallelization—plus when over-decomposition becomes over-engineering.
Seahaven: An Open-Source Framework for Building RL Training Environments for LLM Agents
Seahaven is an MIT-licensed, OpenEnv-compatible Python framework for building reproducible RL training environments for LLM agents, featuring stateful tools, millisecond resets, and verifiable rewards.
TwinCheck: Fixing Tool Agent Errors with "Negative Twin" Counterfactual Verification
TwinCheck uses evidence-grounded "negative twin" counterfactual verification to fix tool agent errors at inference time, boosting success rates from 45.3% to 58.5% on BFCL V4 with zero regressions.
Glyph: A Multi-Strategy Agent System for Automated Enterprise Data Catalog Annotation
Glyph is a multi-strategy LLM agent system for enterprise data catalogs that automates column description generation and sensitivity ontology tagging, grounding outputs in pipeline source code to improve accuracy.
