AI agents rarely crash, they fail quietly, research explains why

Unite.AI examines why LLM agents seldom throw errors, instead returning confident wrong answers while every request succeeds. It cites research showing prompt formatting shifted LLaMA-2-13B accuracy by up to 76 points, and a GPT-4o agent solved 61% of τ-bench tasks once but only 25% eight times in a row.

• Prompt formatting changed LLaMA-2-13B accuracy by up to 76 points
• GPT-4o agent solved 61% of τ-bench tasks once, 25% eight times in a row
• Analysis of 1,600+ traces found 14 multi-agent failure modes (MAST)
• OpenTelemetry is still drafting conventions for generative AI spans