Your Trading Agent Can't Actually Trade, and Your

March 6, 2026

TL;DR

• TraderBench tests 13 AI models on adversarial trading tasks. 8 of 13 use fixed, non-adaptive strategies — extended thinking helps knowledge retrieval but has zero impact on actual trading decisions.
• ETH Zurich finds that LLM-generated AGENTS.md context files decrease coding agent success rates by up to 2% and raise inference costs by over 20%. Human-written files offer only marginal gains.
• NVIDIA argues that small language models — not frontier LLMs — are the right fit for most agentic tasks. With GTC 2026 ten days away, the position paper reads as a strategic preview.
• The Commerce Department's draft AI chip export rules are clashing with the White House, creating regulatory uncertainty across the AI supply chain.