Table of contents

• The goal
• The journalist agent
• How the pieces connect
• Controller and runtime
• AgentTask CRD
• OpenRouter for API auth
• Permissions with allowedTools
• Webhook: trigger on demand
• Helm chart and deployment

The goal

I want AI agents that run autonomously on a cron schedule or in response to events, managed by Kubernetes.

The demo task is a daily AI news digest. Every morning at 8am, an agent searches the web for yesterday's AI news, writes a digest to the wiki, and commits. It's also triggerable on demand via webhook.

K8s handles the scheduling. A custom controller watches a CRD and creates Jobs. The agent runs in a container, writes to a shared volume, and the controller tracks the result. Agents commit locally but don't push. I review and push manually.

The journalist agent

I wrote a new Claude Code agent for this: journalist. Sonnet, not Opus. This runs daily, and the task is search and summarize. Sonnet is cheaper and good enough.

The agent searches for yesterday's AI news, writes a digest to wiki/journal/YYYY-MM-DD/ai-news.md with source URLs for every item, and commits. It determines the date itself via date.

Early versions used WebSearch, which burned through context with HTML scraping. I built a small GNews MCP server that wraps the GNews API. Structured JSON responses, fewer tokens. The agent runs six parallel queries (OpenAI, Anthropic, Google AI, NVIDIA AI, AI startup, plus tech headlines) and deduplicates the results.

I also wrote a Discord MCP server so the agent posts its digest to the #news channel in my Discord bot server.

How the pieces connect

graph TD
    AT[daily-ai-news AgentTask] -->|watched by| CTRL[Controller]
    CTRL -->|8am cron| JOB[K8s Job]
    JOB -->|init| GIT[git fetch + reset]
    GIT -->|then| AGENT[Journalist Agent]
    AGENT -->|writes| WIKI[wiki/journal/YYYY-MM-DD/ai-news.md]
    AGENT -->|commits| PVC[Shared PVC]
    AGENT -->|posts via MCP| DISCORD[Discord #news]
    WH[Webhook :8080] -->|creates| AT

The controller is a Go binary running as a Deployment. It watches AgentTask custom resources in the ai-agents namespace.

Each Job has two containers:

1. Init container (alpine/git): fetches the repo, resets to the configured branch, then chowns the workspace to uid 1000 (the agent user).
2. Main container (ai-agent-runtime): writes MCP config, then runs claude --mcp-config ... --agent journalist -p "..." --allowedTools ...

Controller and runtime

The controller is a Go reconcile loop that lists all AgentTasks every 30 seconds. For scheduled tasks, it compares lastRunTime against the cron expression and creates a Job when due. For manual and webhook tasks, it creates a Job immediately when phase is Pending.

The runtime image is an Alpine container with Claude Code and OpenCode baked in (infra/ai-agents/ai-agent-runtime/Dockerfile). Runs as uid 1000.

AgentTask CRD

The custom resource definition gives kubectl get agenttasks for free. Here's the daily news task:

apiVersion: agents.kyle.pericak.com/v1alpha1
kind: AgentTask
metadata:
  name: daily-ai-news
spec:
  agent: journalist
  runtime: claude
  prompt: >-
    Search for yesterday's most notable AI news...
  schedule: "0 8 * * *"
  trigger: scheduled
  readOnly: false
  allowedTools: >-
    WebSearch,WebFetch,Read,Glob,Grep,Write,
    Bash(mkdir *),Bash(git add *),
    Bash(git commit *),Bash(date *)

The controller splits allowedTools on commas and passes each one as a separate --allowedTools flag to Claude Code. This is how you run Claude Code autonomously without --dangerously-skip-permissions.

Field	Purpose
agent	Name from .claude/agents/
runtime	claude or opencode
prompt	The -p argument
schedule	Cron expression (empty = one-shot)
trigger	manual, scheduled, or webhook
readOnly	Whether this agent modifies the repo
allowedTools	Comma-separated tool permissions

OpenRouter for API auth

Claude Code normally uses ANTHROPIC_API_KEY to talk to Anthropic's API directly. In a container, I route it through OpenRouter instead. One API key for all model providers, unified billing.

The Helm chart's Secret injects these env vars into every agent container:

ANTHROPIC_BASE_URL: https://openrouter.ai/api
ANTHROPIC_AUTH_TOKEN: <openrouter-key>
ANTHROPIC_API_KEY: ""
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: "1"

ANTHROPIC_API_KEY must be empty or Claude Code tries to use it and the auth conflicts.

Permissions with allowedTools

Claude Code prompts for permission before using tools. In a container with no TTY, it just hangs. The fix is --allowedTools, which auto-approves specific tools.

Each tool needs its own flag. Comma-separated doesn't work:

# wrong
claude -p "..." --allowedTools "WebSearch,Write,Read"

# right
claude -p "..." \
  --allowedTools "WebSearch" \
  --allowedTools "Write" \
  --allowedTools "Read"

Tools with specifiers use parentheses for scoping:

--allowedTools 'Bash(mkdir *)'    # only mkdir
--allowedTools 'Bash(git add *)'  # only git add
--allowedTools 'Write'            # any file

I tried scoping Write to a specific path with Write(apps/blog/blog/markdown/wiki/journal/**) but it didn't match inside the container. The pattern works locally but fails when the working directory is /workspace/repo. Unrestricted Write works. The agent definition and Bash scoping still constrain what gets written.

Webhook: trigger on demand

The controller exposes :8080/webhook for on-demand runs. Requires a bearer token via AI_WEBHOOK_TOKEN env var (skipped if empty for local dev):

curl -X POST http://localhost:8080/webhook \
  -H "Authorization: Bearer $AI_WEBHOOK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": "journalist",
    "prompt": "Search for today AI news...",
    "allowedTools": "WebSearch,WebFetch,Read,Write,..."
  }'

The handler creates an AgentTask CR with the allowedTools, and the reconcile loop picks it up.

Helm chart and deployment

The Helm chart packages everything: CRD, controller Deployment, ServiceAccount, RBAC, Secrets, and PVC. hostPath volume, single node.

helm install agent-controller ./helm \
  -n ai-agents --create-namespace \
  -f values-override.yaml

The OpenRouter key has special characters that --set mangles. Use a values file instead:

secrets:
  openrouterApiKey: "<your-openrouter-key>"
  discordBotToken: "<your-discord-bot-token>"
  discordGuildId: "<your-guild-id>"
repo:
  branch: kyle/blog-k8s-autonomous-agents-mvp