AI Root Cause Analysis Shifts from Model Reasoning to Context Engineering

A growing number of observability engineers believe that the reasoning capability of large language models is no longer the bottleneck for AI-assisted root cause analysis (RCA); the more pressing issue is the processing pipeline that decides which data is passed to the model. Most AI RCA work falls into two categories: agent-based designs that give the model tools to investigate autonomously, and deterministic designs that pre-establish correlations between signals and provide a pre-packaged context. Coroot's research, led by engineer Nikolay Sivko, splits the process into two tasks: reasoning over presented data and the 'data collection framework' that decides what data goes to the model. Coroot's pipeline correlates signals into findings and passes them to the model as a single focused context, avoiding agent loops, allowing errors to be attributed to the model itself. Sivko built a scenario using Chaos Mesh NetworkChaos to inject latency between a directory service and its Postgres database, causing slow queries and 502 errors, with misleading signals. He tested eleven models with the same prompt (~9800 tokens), asking each to provide root cause, causal chain, and immediate fix. Frontier models Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro passed, identifying the experiment and the need to delete it and its cron job. Among smaller models, Gemma 4 31B was the only self-hosted model to identify the root cause, while larger Qwen3.6 35B and Qwen3 Coder Next failed. Agent-based approaches have advantages, such as catching unexpected signals, but they suffer from debuggability issues in production, as seen in ZenML and Incident.io cases. Many practitioners now prefer deterministic workflows with a narrow LLM step due to reliability and lower token costs. Sivko notes that a single short call to frontier models costs only a few cents since correlation analysis is done before the call; he believes the reasoning part is 'essentially solved' and current focus is on preparing the right and compact context. This aligns with the broader trend of context engineering, as guidance from Anthropic, LangChain, and Mezmo emphasizes the importance of a minimal, high-signal context set for LLM-based reasoning and observability.