Shepherd — Autonomous SRE Multi-Agent Investigation

An autonomous Site Reliability Engineering (SRE) platform designed to investigate production incidents, query cluster state, and generate verified root-cause analyses without requiring human-in-the-loop manual debugging.

Two-Phase Investigation State Machine

Built on LangGraph state machines, Shepherd enforces a strict separation of concerns among specialized subagents:

1. Evidence Gathering Phase: Specialists (Kubernetes, logs, networking, metrics) query read-only diagnostic tools through domain-scoped MCP servers with strict boundary enforcement.
2. Correlation & Synthesis Phase: Diagnostic tools are withheld while the synthesis engine cross-correlates gathered evidence, preventing models from hallucinating narratives and cherry-picking supporting tool outputs.

Key Capabilities

• Deterministic Evaluation Gates: Validates incident hypotheses against deterministic evaluation criteria; automatically triggers targeted deep-dive loops if confidence is insufficient.
• Multi-Tier Model Fallbacks: Resilience chains that fall back across alternative LLM providers during latency spikes or upstream outages.
• RAG-Powered Runbooks: Contextual retrieval of organizational troubleshooting guides and operational playbooks.
• Automated Validation: Validated end-to-end against live incident scenarios injected into ephemeral KinD Kubernetes clusters in CI.

Technology Stack

• Runtime & Orchestration: Python 3.11+, LangGraph, FastAPI
• Tool Protocols: Model Context Protocol (MCP) with domain-isolated servers
• Intelligence: RAG over Runbooks, Multi-Tier Model Fallback Chains, Deterministic Evaluation Gates
• Infrastructure: Kubernetes (KinD), Docker, GitHub Actions