A growing number of observability engineers now believe that large language model reasoning ability is no longer the bottleneck in AI-assisted root cause analysis. The harder problem, they argue, sits in the pipeline that decides what data reaches the model. This shift in thinking, documented in a new InfoQ article published July 25, 2026, points to a practical takeaway: effort spent preparing context may pay off more than reaching for a larger model.
The article, written by Mark Silvester, examines how the industry is splitting into two camps: agent-based designs, where the model is handed tools and allowed to investigate and choose telemetry as it reasons, and deterministic designs, where signals are correlated up front and the model receives a single prepared context. Coroot, an observability vendor, has conducted original research that reflects an industry shift toward the deterministic camp. Dynatrace's Davis AI similarly leans on deterministic, topology-based causal analysis, traversing a real-time dependency map.
The Two Jobs of AI-Assisted RCA
Coroot's original research, designed and run by engineer Nikolay Sivko, split AI-assisted root cause analysis into two distinct jobs: reasoning and harness. Sivko argued that "can AI do RCA?" is the wrong question, and that the two jobs need to be measured apart. Coroot's pipeline correlates signals into findings and passes them to the model in one focused context with no agent loop. This means a wrong answer can be blamed on the model rather than on missing evidence.
The two approaches break in different ways, making it hard to tell whether a failed diagnosis stems from weak reasoning or a wrong evidence harness. Agent-based designs hand the model tools and let it investigate, choosing telemetry as it reasons. Deterministic designs correlate signals up front and hand the model a single prepared context. Coroot's work reflects a clear preference for the latter.
Experiment: Eleven Models, One Scenario
Sivko designed an experiment testing 11 models on a single scenario. Chaos Mesh, a chaos engineering tool, was used to inject network delay between a catalogue service and a Postgres database. The delay slowed queries and pushed the front-end into 502 errors. The context deliberately included misleading signals: query timings inflated by network round-trip time.
The same prompt of approximately 9,800 tokens was run against all 11 models. Each model was asked for the root cause, a cause-and-effect chain, and an immediate fix. The closed frontier models—Claude Opus 4.8 from Anthropic, GPT-5.5 from OpenAI, and Gemini 3.1 Pro from Google—all passed. They named the experiment and flagged the need to delete it and its schedule.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Larger open-weight models mostly kept pace. Gemma 4 31B was the only self-hostable model to identify the root cause. Qwen3.6 35B and Qwen3 Coder Next both missed it. The findings do not settle the debate, but they highlight a clear gap between closed frontier models and open-weight alternatives.
The Trade-Off: Flexibility vs. Operability
Agent-based approaches keep a genuine advantage: the model can pull signals that a fixed pipeline never anticipated, which is important for novel incidents. But that flexibility is paid for in operability. Accounts from ZenML and Incident.io describe multi-agent LLM investigations as notoriously hard to debug in production. A failed agent run leaves no clean stack trace, only unpredictable prompt interactions and emergent coordination.
Engineers on Reddit report scrapping fully agentic designs for mostly deterministic workflows with a narrow LLM step, citing better reliability and lower token cost. Deterministic pipelines trade flexibility for repeatability and cleaner evaluation. The industry increasingly treats the harness, not the model, as the hardest part of AI-assisted root cause analysis.
Sivko noted that a single short call runs only a few cents, even on frontier models, because the correlation work happens before the model call. He judged the reasoning part of AI-assisted root cause analysis "basically solved." His conclusion points the next engineering effort at the harness rather than the model. "Preparing the right, compact context for the model before you call it," Sivko said, is where the real work lies.
Context Engineering Emerges as Core Discipline
This framing echoes a wider push around context engineering. Guidance from Anthropic, LangChain, and Mezmo converges on curating the smallest set of high-signal, compact context as a core discipline. The idea is that effort spent preparing context may pay off more than reaching for a larger model.
The article, published on InfoQ as a 3-minute read, presents Coroot's research as evidence of an industry shift. It contrasts the two approaches and highlights their trade-offs. The conclusion is clear: the harness—the pipeline that decides what data reaches the model—is now the harder problem than model reasoning. Context engineering is emerging as a core discipline for making LLM-based observability reliable.