Agentic operations · Amazon Bedrock · AgentCore · CloudWatch · OpenTelemetry
Infrastructure Monitoring Copilot
An AWS operations agent built with Amazon Bedrock and AgentCore that correlates telemetry, investigates likely causes and proposes remediation while engineers retain control.
Architecture: Amazon Bedrock + Amazon Bedrock AgentCore
The challenge
Cloud incidents generate noisy metrics, logs, traces, deployment events and alerts across services. Engineers need fast context, but an autonomous agent must not make unreviewed production changes.
My end-to-end ownership
I independently delivered the complete platform from discovery to handover. I owned product and workflow design, solution architecture, LangGraph orchestration, AWS integrations, telemetry correlation, runbook retrieval, approval controls, frontend and backend engineering, infrastructure, testing, observability, deployment and documentation.
- Discovery and requirements
- Product and solution design
- Architecture and engineering
- AI and data workflows
- Cloud, security and CI/CD
- Testing, observability and handover
System design
The copilot correlates related signals, summarises impact, proposes likely causes, retrieves relevant runbooks and builds a remediation plan. Suggested actions can be simulated before an authorised engineer approves execution.
AWS operations architecture
The agent investigates and recommends. Deterministic AWS services continue to own alarms, workflow state, approvals and production changes.
Metrics, logs, traces, deployment events and alarms arrive with correlated service and request context.
A LangGraph-style agent builds hypotheses, asks for missing evidence and produces a structured incident plan.
Retrieves current runbooks, architecture records and previous incident evidence with source attribution.
Exposes narrow CloudWatch and Systems Manager operations using scoped identity and deterministic tool policy.
Owns incident state, retries, approval waits, escalation and safe recovery outside the model loop.
Measures diagnosis quality, tool accuracy, time to evidence, unsafe-action attempts, latency and cost.
Remediation is first simulated or previewed. AWS Systems Manager Automation or another approved runbook executes only after the required human authorization and policy checks.
Production controls
Evidence-linked reasoning, confidence scoring, explicit tool permissions, human approval, idempotent actions, audit logs, OpenTelemetry traces and fallback escalation keep engineers in control.