Modernizing Incident Response with LLMs, RAG, and the MCP
summary
A large SRE team replaces fragmented scripts with an LLM-driven workflow that uses the Model Context Protocol (MCP) to inject real-time logs, configs and tickets into prompts while preserving access controls; a domain-specific retrieval layer built on OpenSearch + Bedrock Titan embeddings provides semantic search over incidents and playbooks; image-rendered Grafana panels outperform raw CSV for anomaly reasoning; a Promptu benchmark plus thumbs-up human feedback tracks pass rates and drives prompt/versioning, yielding a measurable drop in mean time to resolution and gated human approval for any destructive action.
AI-written summary · about
chapters (14)
- 0:00Modernizing incident response
- 2:27Scripts to agentic cognition
- 6:19Institutional knowledge loss
- 8:04Single-prompt resolution
- 10:08Guardrails for collaboration
- 11:55Unified MCP tools
- 14:29Prompting and indexing
- 19:35Task decomposition
- 21:40Evaluating with Promptfoo
- 24:49Feedback loops and metrics
- 28:43Visualizing cooperation
- 31:55Evaluation flywheel
- 33:39Cognitive reliability
- 34:42Key takeaways
related talks
36mSRE'ing the SRE AI Agents
27mAccelerating Incident Response With OpenTelemetry and MCP: Building...
34mAgentic Incident Response
17mIncident Response @ Scale-Building a Next Generation SOC
29mComposing the Response: Building an Incident Pipeline from Scratch
40mWhen Incidents Fix Themselves: AI SRE in action
25mLLM Copilots for SRE: Incident Response & Auto-Remediation
24mAI SRE and Agent Handoffs: incident.io's Collaboration Model