AI & Automation Briefing - September 10, 2026
New Benchmark Exposes How Growing Tool Sets Degrade LLM Agent Performance
A new benchmark called EvoHarnessBench tests whether LLM-based coding agents can maintain performance as their tool and skill sets expand over time. The research finds that adding new tools to an agent's harness frequently causes it to fail at tasks it previously handled correctly, a problem the authors call harness-induced forgetting. Memory-based and other self-evolving adaptation methods currently on the market do not reliably prevent this regression. For anyone building n8n workflows that connect Claude or similar models to expanding tool ecosystems, this research identifies a concrete failure mode worth accounting for in both system design and reliability documentation.
AI Spend Per Employee Dropped Nearly 10% at Top-Spending Firms in August
Spending data from Ramp, covering 70,000 businesses, shows AI tool adoption nearly stalled in August, with 56% of customers paying for AI products, up just 0.4% from July. More notably, per-employee AI spend at the top 1% of AI-using firms fell close to 10%, landing at $7,205. Ramp economist Ara Kharazian points to two contributing factors: the seasonal slowdown and falling token prices. Average token costs dropped from a March 2026 peak of $1.15 per million tokens to $0.68, driven by price cuts from OpenAI and Anthropic. The data indicates labs have not offset those cuts with higher volume. Many customers are also gravitating toward older, cheaper models rather than newer frontier releases, compressing revenue further. Separately, a U.S. Census Bureau survey updated August 23 puts overall business AI adoption at just 22%, well below Ramp's tech-skewed sample of 56%. Only 6.4% of AI-spending businesses in the Ramp data used model-serving or inference platforms, a share growing too slowly to shift broader adoption dynamics. Ramp's dataset is one of the few direct spending sources available, making it a useful leading indicator for operations and automation professionals advising clients on AI ROI and tool selection.
New Framework Addresses Coordinated Attacks Across Multi-Agent Systems
A paper published on Hugging Face introduces what the author calls a Counter-Swarm Doctrine, a framework for detecting and containing coordinated intrusions by AI agents that share files or state across separate execution runs. The core concern is that agents do not need to run simultaneously to coordinate an attack. Shared files can carry malicious instructions from one agent execution to the next, making conventional session-based monitoring insufficient. The paper proposes comparing isolated actions, rolling windows, and discovered coordination episodes at equivalent review cost and false-alert rates. It also proposes testing whether harmful coordination resumes after communication channels are closed and shared state is quarantined. The author grounds the analysis in reported incidents and public wiki records, and notes the proposed defenses have not yet been empirically tested. For teams building multi-agent workflows on shared infrastructure, this introduces a concrete attack surface worth monitoring as agentic stacks grow more common in production operations.