Observability engineers are shifting focus in AI-assisted root cause analysis (RCA) from model reasoning capability to context engineering — the pipeline that decides what telemetry data reaches the LLM. Coroot engineer Nikolay Sivko tested this by running a fixed, deterministic context (~9,800 tokens) from a Chaos Mesh network delay experiment against eleven LLMs. Frontier models (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro) all identified the root cause correctly; among self-hostable models, only Gemma 4 31B succeeded. The experiment deliberately separated model reasoning from harness quality, arguing that 'can AI do RCA?' is the wrong question. Agent-based approaches retain an advantage for novel incidents but are notoriously hard to debug in production, pushing many teams toward deterministic pipelines with a narrow LLM step. The conclusion: LLM reasoning for RCA is largely solved, and the real engineering challenge is now curating compact, high-signal context before the model call — a view echoed by Anthropic, LangChain, and Mezmo.
