Causely Causal AI Layer Improves SRE Agent Reliability

A recent preprint on arXiv presents Causely, a causal intelligence layer aimed at improving AI agents within Site Reliability Engineering (SRE) processes. This system mitigates the semantic-interpretation burden that arises when agents interpret environmental data from raw observability telemetry during queries, which can lead to increased token consumption, delays, and unreliable inferences. Causely offers a structured view of environment topology, attribute dependencies, and causal links based on an ontological framework. By converting raw telemetry into a dynamic, queryable model, it equips AI agents with the necessary semantic and causal insights for effective issue diagnosis, impact evaluation, and safe actions in production. The researchers tested Causely in a controlled benchmark study using a 24-microservice OpenTelemetry demo application with injected faults, comparing four agent setups: Claude Code, OpenAI Codex, HolmesGPT with Sonnet, and HolmesGPT with Gemini backends. The findings revealed that agents utilizing Causely outperformed baseline configurations in both accuracy and latency for fault diagnosis and impact assessment. The paper can be found on arXiv with the identifier 2605.18327.