Guides

Incident & SRE agents

On alert → query metrics/logs/deploys → hypothesize RCA → draft PR stub. agent_id: incident.

Scenario

Two agents on the same Kafka-lag page: one blamed CPU, one found consumer lag. TraceForge side-by-side plus FineForge eval cases teach which tool graph is correct.

What Inferix gives you

TraceForge tool graphs for metrics → logs → deploys → draft PR
LensAI cost/latency of long investigation traces
RouteIQ: classify intent cheaply, escalate ambiguous RCA to general-llm
DriftWatch when retrieval/poisoned runbooks hurt RCA quality

Tools & intents

Tools	Intents
pager.get_alert, metrics.query, logs.search, git.recent_deploys, k8s.get_events, github.open_draft_pr	latency_regression, error_budget_burn, kafka_lag, oom_kill, bad_deploy, dependency_outage

Setup

routing:
  rules:
    - name: incident-triage
      when: { agent: incident, intent: [oom_kill, kafka_lag] }
      model: owned/general-llm
    - name: incident-ambiguous
      when: { agent: incident, confidence: low }
      model: provider/strong

# Require git.recent_deploys before blaming infra-only causes
# Golden: kafka_lag cases must include consumer lag tool span

Watch

N+1 metrics.query loops (T06)
Missing deploy tool on bad_deploy intents (T01)
Exclusive time on logs.search (bottleneck)
Poison retrieval chaos → DriftWatch score

Failure modes

Drop deploy tool → wrong RCA. Inject latency on logs.search → incomplete hypotheses. Force expensive model → LensAI cost spike without quality gain. See TraceForge.

Invoice reconciliation and ERP schema drift.

Finance agents →