AI & Automation Briefing - July 19, 2026
GRASP Introduces Granularity-Aware Search Decisions for Agentic RAG Systems
A research paper published July 11, 2026 introduces GRASP, a search policy designed for agentic retrieval-augmented generation that controls retrieval granularity as language models iteratively reason and query external sources. Standard agentic RAG systems retrieve information without adjusting the scope or detail level of each search step. GRASP addresses this by giving the agent a policy for deciding when to retrieve broadly versus narrowly, improving how multi-step reasoning tasks consume retrieved context. For operations teams building AI workflows that pull from internal knowledge bases or external data sources, granularity-aware retrieval is a meaningful step toward more accurate and efficient agent behavior in production pipelines.
LongStraw Enables Million-Token RL Training Without Expanding GPU Budget
Researchers have released LongStraw, an execution framework that makes reinforcement learning post-training feasible on sequences exceeding 2 million tokens without requiring additional GPU memory. The system evaluates the shared prompt once without storing the full computation graph, keeps only the model state needed for subsequent tokens, and replays individual response branches one at a time under automatic differentiation. This approach reduces live memory to a single response branch at any given moment, trading some extra compute time for significantly lower GPU overhead. LongStraw has been validated on two different model architectures: the hybrid recurrent and full-attention Qwen3.6-27B, and the compressed-attention mixture-of-experts GLM-5.2. For operations teams running large-scale AI workflows on fixed infrastructure, the technique points toward a path for fine-tuning long-context models without upgrading hardware.
Patreon Moves from Asking to Actively Blocking AI Training Bots
Patreon has partnered with Cloudflare to actively block AI training crawlers from scraping creator content, replacing its previous approach of using robots.txt files to request compliance. The move was prompted by evidence that scrapers were ignoring those requests. During testing, weekly access attempts from individual AI training bots dropped from thousands to zero. Patreon says the shift was also driven by new platform features like a redesigned Home Feed and a short-form post format called Quips, which exposed more content to crawlers than its paywalled content previously did. The company will continue allowing bots that index pages for search and discovery purposes. Cloudflare has been expanding its suite of AI crawl management tools, including a Pay Per Crawl marketplace and a recent policy change that blocks mixed-use crawlers by default on ad-supported pages.