AI & Automation Briefing - August 19, 2026
StateM Hits 95.3% Accuracy on Terminal-Bench 2.1 Without Model Fine-Tuning
A new paper introduces StateM, a runtime harness that improves AI agent reliability on complex terminal tasks by adding durable state tracking, recoverable runbooks, and procedural controls around the model rather than inside it. On Terminal-Bench 2.1, StateM raised GPT-5.5 accuracy to 92.1% against an 83.1% baseline and edged out GPT-5.6 Sol Ultra at 91.9%. The same runbook transferred without changes to GPT-5.6, where StateM reached 95.3% raw accuracy across 445 trials and completed all 89 tasks successfully at least once. The approach, which the authors call harness scaling, is model-agnostic and requires no fine-tuning, making it directly applicable to long-horizon agent workflows built on tools like n8n and Claude.
Linear Data Shows AI Adoption Doubling Across Every Software Team Function in Six Months
Linear published behavioral data from 127,000 paid users showing AI feature adoption more than doubled across every job function between January and June 2026. Product managers saw the largest gain, rising from 12% to 34% active on AI features. Go-to-market roles, the furthest from the codebase, climbed from 5% to 18%. The most notable executive finding: CEOs at companies with more than 200 employees jumped from 9% to 36%, the largest single increase in the report. Linear frames this as a picture of AI-assisted product development across the full workflow, from issue creation to pull request, not just code output metrics. The data covers activity inside Linear only and does not represent broader market adoption.
No Single Memory Type Wins: Research Maps Trade-offs Across LLM Agent Memory Substrates
A new paper from a team of researchers evaluates eight categories of memory substrate for LLM agents, including dense and sparse retrieval indexes, structured stores, hierarchical memory, refinement-based memory, parametric updates, and activation-compatible context. Testing across three backbone models and four benchmark suites using 26 performance and efficiency metrics, the study finds that no single memory type consistently outperforms the others. Performance depends on task type, context length, and operating conditions. The authors conclude that adaptive substrate routing, selecting the right memory type based on situational factors, is the practical path forward for building reliable long-horizon agents. For teams connecting models like Claude to workflow tools such as n8n, the findings offer a concrete framework for deciding how to persist and retrieve context across automated processes.