AgentOps

AI Agents as Kubernetes Workloads

Define agents, tools, memory, and channels as Custom Resources. The operator handles the rest.

Documentation

What is AgentOps?

A Kubernetes-native platform for running AI agents in production. No wrapper scripts. No Docker-in-Docker hacks. No sidecar orchestrators glued together with YAML.

Agents are first-class Kubernetes workloads with their own CRDs, controllers, memory, tool access, and observability. kubectl apply an Agent CR and get a running, observable, delegating AI agent with persistent memory and streaming output. Built on the Charm Fantasy SDK. Pure Go. No Python runtime.

Principle 01

Kubernetes-Native

Agents are CRDs, not containers wrapping scripts. Reconciliation loops, not cron jobs. The control plane, not a sidecar.

Principle 02

Zero Abstractions

No framework lock-in, no SDKs to learn. Pure Kubernetes primitives. If you know kubectl, you know AgentOps.

Principle 03

Observable by Default

Every tool call, memory read, and delegation traced end-to-end with OpenTelemetry. You see everything.

Principle 04

Production-Grade

Single static Go binaries. Predictable resources. No cold-start surprises. Built for SRE teams who run real infrastructure.

Platform Capabilities

Every component designed for production. Every integration first-party.

Three-Layer Memory

Working memory (token-budget trimmed), short-term (deterministic session summaries), long-term (user-curated). Context injection is BM25 relevance-ranked via FTS5. No embedding models, no vector DB.

Learn more →

Agent Delegation

Agents spawn sub-agents as Kubernetes Jobs with independent tools, memory, and resources. Fan-out, fan-in. Zero-polling Watch for result aggregation. Concurrency at the CRD level.

Learn more →

MCP Tools as OCI

Tool servers are compiled Go binaries with MCP stdio transport. Package as OCI artifacts, push to any registry, reference in your Agent CR. Pulled at reconcile time by init containers.

Learn more →

Real-Time Console

SolidJS PWA with Go BFF. FEP over Server-Sent Events for live streaming. 12 specialized tool card renderers. Tempo trace integration. Memory management panel.

Learn more →

OTEL Observability

Every turn, tool call, memory read/write, and delegation traced end-to-end with OpenTelemetry. Per-observation injection audit trails. Traces to Tempo, metrics to Prometheus.

Learn more →

Go-Native Runtime

Built on the Charm Fantasy SDK. Single static binary per agent pod. Fast cold starts, predictable resources. Handles memory injection, tool dispatch, and FEP streaming.

Learn more →

Define an Agent

An Agent CR is everything Kubernetes needs: model, system prompt, tools, memory, delegation rules, and resource limits.

apiVersion: agents.agentops.io/v1alpha1
kind: Agent
metadata:
  name: site-reliability
  namespace: agents
spec:
  model:
    provider: anthropic
    name: claude-sonnet-4-20250514
  systemPrompt: |
    You are an SRE agent responsible for the production cluster.
    Investigate alerts, correlate with recent deployments, and
    propose remediation. Delegate deep-dives to specialist agents.
  toolRefs:
    - name: kubectl-tool
      registry: ghcr.io/samyn92/agent-tools/kubectl:v0.3.0
    - name: prometheus-tool
      registry: ghcr.io/samyn92/agent-tools/prometheus:v0.2.1
  memory:
    workingMemory: {}
    shortTerm:
      enabled: true
    longTerm:
      enabled: true
  delegation:
    maxConcurrent: 3
    agents:
      - name: log-analyzer
      - name: deployment-checker
  resources:
    requests:
      memory: "256Mi"
      cpu: "250m"
    limits:
      memory: "512Mi"
      cpu: "500m"