Faultline

Faultline is an open-source personal project exploring what a Kubernetes-native AI agent runtime looks like when built from first principles with systems engineering rigor.

Core Architecture

Rather than treating AI agents as external scripts or opaque black-box containers, Faultline defines agents and their model configurations as native Kubernetes Custom Resource Definitions (CRDs):

1. Custom Resources (FaultlineAgent & FaultlineLLM): Allows declarative orchestration of agent state, memory limits, tool access, and model parameters via standard Kubernetes manifests.
2. Reconciler Loop: A Go operator built with controller-runtime that continuously reconciles declared agent specifications with running pods and tool connectors.
3. Allowlist-Only MCP Tool Security: Implements a strict zero-trust boundary where agents cannot invoke arbitrary endpoints. Tools must be explicitly registered and bound to the agent specification.
4. OAuth Dynamic Client Registration (DCR): Implements automated, ephemeral OAuth registration for remote HTTP MCP servers, eliminating hardcoded client credentials.
5. Write-Only Secret Handling: Integrates with Kubernetes Secrets ensuring API keys and auth tokens never round-trip or leak into client-side UI telemetry.

Philosophy

At enterprise scale, AI agents cannot exist as unmonitored scripts. Faultline proves that agents can be managed with the same declarative reliability, lifecycle control, and security boundaries that engineers expect from container orchestration.