My OpenClaw Setup in 2026: The Agent Stack I Run 24/7

I got a question in a DM last week that I think is worth answering publicly: "What does your actual OpenClaw setup look like in 2026?"

So here's what I run, what I changed recently, and what I'd tell myself if I was setting it up from scratch today.

The Core Setup

One primary agent (my main session) that handles:

• Daily research and digest generation
• Task management and reminders
• Research synthesis and writing
• Telegram messaging
• File operations and content management

One research subagent that runs the daily AI trend cron job. Spawned per-task for specific investigations.

One code agent for larger refactoring tasks, using Claude Code for the heavy lifting when OpenClaw's session system isn't the right tool.

What Changed in the Last 30 Days

Upgraded to 2026.5.16-beta.1 for the ambientTurns: "room_event" config. My group chat setup stopped being annoying. That alone was worth the upgrade.

Enabled the file-transfer plugin with the default-deny policy on paired nodes. Had a moment where I was testing something and the symlink traversal protection caught an edge case I didn't know existed. Better to have it and not need it.

Set up the plugin CLI for validating third-party plugins before installing. I was burned by a bad plugin manifest that caused a cryptic failure last month. The openclaw plugins validate check now runs before any new plugin install.

The Config That Actually Matters

{
  "messages": {
    "groupChat": {
      "ambientTurns": "room_event"
    }
  },
  "plugins": {
    "entries": {
      "file-transfer": {
        "enabled": true,
        "defaultDeny": true
      }
    }
  },
  "agents": {
    "defaults": {
      "model": "claude-sonnet-4-20250514",
      "thinking": "medium"
    }
  }
}

The thinking: "medium" is important. For research and writing tasks, medium thinking gives me better output quality without the latency cost of high thinking. For simple message handling, I let it default to low.

The Cron Jobs That Run Without Me

1. Daily AI research — every morning at 7:04 AM ET. Web search, research synthesis, article angles, file creation, Telegram notification. This runs in an isolated session with a model fallback chain.

2. Heartbeat check — every 30 minutes. Checks if the gateway is healthy, restarts if needed, reports status to me if something went wrong.

3. Weekly config review — every Sunday. Checks for config drift, validates plugin list against what I expect, flags anything that changed unexpectedly.

The Tool Chain

• Telegram for messaging (primary notification channel)
• Obsidian via the OpenClaw skill for notes and knowledge management
• Git for content versioning (all articles and research in git-tracked directories)
• Browser for web research and content publishing
• File operations via the bundled file-transfer plugin

What I'd Tell Myself Starting Fresh

The same things I tell everyone who asks:

1. Start with one agent, not many. Get the use cases solid before adding subagents.

2. Treat the config as code. Version it, review changes, don't edit it while the agent is running.

3. The session system is the product. Understand session targets, session state, and how to route work to the right session before you worry about models.

4. Upgrade regularly but test first. The 2026.5.x series has been solid. I run beta on a non-production instance for a week before upgrading the main setup.

5. The skills system is underrated. Most users don't use it enough. The Obsidian skill alone has replaced three separate workflows I used to do manually.

---

The setup: OpenClaw 2026.5.16-beta.1, Mac Mini M4 Pro, 24GB unified memory, running continuously since late 2024. Primary use: content research and writing automation.