A software system spends more than 90% of its lifecycle in the running state, yet Tencent Cloud CloudQ technical lead Yao Binbin told AICon 2026 in Shenzhen that operations intelligence has lagged well behind coding: individual coding throughput has risen roughly two to three times, while keeping systems stable still absorbs large amounts of manual effort. CloudQ's answer is an AI-native SRE agent built on a panoramic running-world graph that fuses infrastructure topology, live observability signals and operational knowledge into one queryable causal structure, and the reported engineering practice spans 8,092 Chinese characters of talk transcript published by InfoQ on 10 October 2026.
"The real bottleneck is not how many tools you have, it is how people understand and make decisions," Yao said, describing why three generations of SRE tooling failed to lift the burden: "tools can put the data in front of you, but reading the data and making the judgment still depends on a human being." Yao framed the gap as a structural one. Coding work is reproducible and testable, while the running world faces constant uncertainty from environment, infrastructure and dependencies, whether a fiber line is cut or a data centre catches fire, so the running side resists full automation far more stubbornly than writing code does.
That read changed because the substrate finally matured. Infrastructure configuration, OpenTelemetry-style metrics, logs and traces, change events and alerts, plus architecture documents, past incidents and tickets had each become reliable in isolation, and frontier models crossed a production-usable threshold while tool calling and harness engineering matured. CloudQ therefore treats the running world as three complementary datasets that answer different questions: infrastructure says what the system looks like, observability says what is happening right now at second or minute granularity, and knowledge says how similar problems were solved before. An ontology layer defines the entities, attributes, relationships and hard constraints, and a semantic layer aligns entities and resolves ambiguity, together producing the graph the agent reasons over.
The operational workflow the talk traces runs from coding back to coding: one alert on a forum service triggered an agent-led architectural diagnosis, drawing the full request chain from entry point through backend to database, checking upstream traffic when the alerted CVM showed high memory, moving into service logs when traffic looked normal, and landing on a stack trace where an unreleased resource leaked the MySQL connection pool. The agent then proposed a temporary recovery of clearing idle connections and raising the connection cap, alongside a long-term code fix. The same graph supports change risk control across three phases, blast-radius mapping one and two hops out before a change, baseline comparison against live metrics during it with blocking when anomalies correlate, and rollback plans plus a summary report after it.
Production hardening is where the architecture gets concrete. A Harness layer accepts input from enterprise IM tools such as WeCom, Feishu and DingTalk as well as a web console, and exposes the whole agent as a skill callable by other agent systems. A SuperAgent orchestration hub handles understanding, planning, execution, context and safety control, sitting above a multi-cloud adapter layer for Tencent Cloud, AWS, Alibaba Cloud and Huawei Cloud, with model services, data analysis and sandboxes underneath. The graph engine itself pins the ReAct loop to a fixed set of nodes rather than letting the model roam, since Yao noted prompts are easily bypassed and execution boundaries are better enforced by rules outside the model.
Scheduled work is treated as an execution plan engine rather than a repeated agent run. Yao said a naive timer approach costs tens of yuan per execution on higher-end models, takes seconds to tens of seconds per model round trip, and still fails non-deterministically; log cleanup or the data-gathering half of a capacity or market analysis can simply be scripted, so the engine decides which steps become scripts and which stay with the model, hands failed steps to the model, and falls back to a full agent retry when the whole task fails. Evaluation runs real-time and offline: sanitised user turns go to a multi-model reviewer scoring results and trajectories across five dimensions, with the next user turn also read as a signal, and offline regression runs twice daily at 08:00 and 20:00 plus on every agent release, blocking deployment when cases fail.
Yao closed by laying out a staged trust model that CloudQ now runs simultaneously: read-only scenarios such as root-cause investigation are opened to autonomous exploration with audit trails, advisory scenarios put the agent in a proposing role while a human authorises the action, and narrow low-risk write operations such as scaling capacity are pre-authorised for unattended execution. That third tier, which Yao called L3, only became viable this year as agent capability improved, and he positioned the overall path, data first, then graph, then harness and evaluation, then limited execution, as a reference construction route for other teams building SRE agents.