AI SRE — multi-agent triage with human gate
supervisor
breaks incident into sub-queries
synthesizes hypotheses
log agent
tool: log search
pattern detection, frequency
grouping by service + trace
metric agent
tool: PromQL / DCGM
correlation windows
anomaly + change-point
runbook agent
tool: vector search
matches to RB-###
proposes safe action
adversarial evaluation (NeuBird pattern)
agents agree → high confidence → present to human
agents disagree → escalate with both hypotheses
filters against hallucinated root causes
human approval gate
safe actions: restart pod, revert deploy, scale within bounds
broad changes always escalate: topology, code, IAM
"set and forget" is overselling