Context Engineering for Agents
Techniques for constructing, pruning, and augmenting context windows to improve LLM agent performance and reliability.
43 papers · 3 months
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
June 2026
Pruning context to recent tool calls improves agent reliability and efficiency
Pruning context to recent tool calls improves reliability and efficiency, comparing four context engineering strategies.
6 further papers
Sept 2026 · Zhejiang University, Apple
Paper-derived rubrics improve AI generation of scientific research plans
Demonstrates deriving prompts and evaluation rubrics from scientific papers for research-plan generation, advancing context augmentation for domain-specific agent tasks.
released code
Sept 2026 · Perplexity AI
Benchmark tests retrieval for agent-written queries across 190 million web pages
Introduces Q2D-Web benchmark for evaluating retrieval in agent-written queries, informing context window construction for agents.
2 further papers
Sept 2026 · Rutgers University
Kernel-managed memory personalizes multiple AI agents with shorter prompts
Shows how kernel-controlled retrieval shapes context windows to reduce prompt length while personalizing agent behavior.
3 further papers
Sept 2026 · Mohamed bin Zayed University of Artificial Intelligence, University of Science and Technology of China
Deep research agents struggle with long, multimodal evidence chains
Benchmark reveals that long, multimodal evidence chains stress context construction and reasoning.
released code
Sept 2026 · University of California San Diego
Coding agents retrieve better by targeting missing evidence sets
Proposes MSS-Complement retrieval that targets missing evidence, directly improving context construction for coding agents.
2 further papers
Sept 2026 · NLPR&MAIS, Institute of Automation, Chinese Academy of Sciences
Bounded visual workspaces improve multimodal agents’ accuracy and efficiency
Bounded visual workspace via ledger storage reduces context length and improves accuracy.
1 further paper
Sept 2026 · Amazon.com
Agentic document QA pays off mainly with stronger vision-language models
Compares static and agentic pipelines for document QA, finding that agentic approaches only become competitive with stronger vision-language models, providing insights for context engineering.
2 further papers
Sept 2026 · University of Southern California
Short natural context can flip decision models to wrong answers
Shows that short, fluent context additions can flip correct decisions, highlighting context fragility.
3 further papers
Sept 2026 · RIKEN
Execution Logs Can Mislead Visual Judges in Video-Generation Agents
Shows that execution logs in the judge context can mislead visual verification, a reliability constraint for context construction.
Sept 2026 · Rutgers University, Nanjing Tech University
LLM pipelines can lose 40.5 accuracy points between stages
Quantifies the decomposition tax across LLM pipeline stages and shows re-grounding later stages with the original problem reduces interface information loss.
2 further papers
Oct 2026 · The University of Hong Kong, Huawei Research
Parallel tile inspection helps models find details in large images
Uses parallel tile inspection to manage visual context, improving detail detection in large images.
1 further paper
Oct 2026 · Carnegie Mellon University, AWS AI Labs
Models Learn to Allocate Context and Scale Inference More Effectively
Introduces Hermes harness for context allocation, improving inference effectiveness.
1 further paper
Oct 2026 · Adobe Research, University of Maryland College Park
Voice agents lose document grounding as conversations and context grow
Shows document grounding degrades with context length in voice agents, motivating better context management.
2 further papers
Oct 2026 · KU Leuven, East China University of Science and Technology
Persona effects in language models are predictable but not portable
investigates whether persona prompt effects generalize across models, finding they are predictable within a model but not portable.
released code
Oct 2026 · Columbia University, Amazon Web Services
General instructions outperform gold answers in on-policy context distillation
Shows simple formatting instructions outperform gold answers for on-policy context distillation, improving context construction for agents.
16 of 41 papers shown