OpenClaw 2.0: what shipped, and what it means for running agents 24/7

What actually shipped

This is the biggest release in OpenClaw's history: 933 contributors (569 of them first-timers) and more than 16,000 pull requests, roughly half of every PR ever merged into the project, packed into one release. The team had shipped 106 releases in 230 days before this, then went quiet for nearly seven weeks to make it hold together.

• A rebuilt browser app: the web experience opens directly into a conversation with your agent, and setup, ongoing work, and live monitoring all happen there instead of in a half-wired dashboard.
• Guided onboarding that uses what you already have: setup scans your machine for existing ChatGPT, Claude, and Codex subscriptions, API keys, and local models (Ollama, LM Studio, llama.cpp), verifies the model actually works before saving it, then hands off to a first conversation.
• Shared cloud sessions, or multiplayer agents: you can bring another person into a live agent session or hand a task over with full context intact.
• Sessions moved to SQLite: sessions and transcripts now live in a database instead of flat files, which is more robust for 24/7 operation and is also the one migration with a real downgrade trap.
• A reliability pass across everything: Gateway startup, credentials, plugin handling, mobile pairing, and messaging channels, with hundreds of fixes in the long tail.
• Fresh OpenAI setups default to GPT-5.6, and local-model setups get a managed llama-server path with a 64K default context.

Sources: OpenClaw 2.0 release notes

Why this is not a minor bump

OpenClaw 2.0 changes how you think about agent lifecycle, state management, and sharing work. If you are running a single agent on a laptop, you probably will not notice. If you are running five agents across three projects with cron jobs, webhooks, and external API calls, you will notice immediately. Shared sessions in particular change what OpenClaw is, from a personal tool into a shared workspace for teams running agents together.

What this means for 24/7 agents

Most people testing OpenClaw run it for an hour, maybe a day, then kill the process when they are done. That is not production. Production is an agent that runs while you sleep, handles failures gracefully or does not, manages state across restarts or loses everything, and costs you money every hour it runs. OpenClaw 2.0 changes the calculus on all four.

• Uptime: the release includes a long reliability pass on Gateway startup, and startup stays responsive while model runtimes are prepared instead of hanging.
• State persistence: the SQLite migration means session state survives restarts as a database, not a pile of files you can half-lose.
• Failure recovery: openclaw doctor is the health-check entry point after any update, and openclaw update runs it automatically. Your verified backup is the rollback path.
• Cost: guided setup detects local models, and verified local models automatically run with a lean tool surface, which matters if you are metering usage.

The upgrade question

Upgrading a running agent is not like upgrading a web app. With a web app you deploy, test, and roll back if it breaks, and downtime is minutes. With an agent you have active conversations in progress, scheduled jobs that cannot be paused mid-run, state files that might not be backward-compatible, and external integrations that expect specific behavior. Break the wrong thing and you have orphaned processes, lost context, and a customer wondering why their agent stopped responding at 2am.

• 2.0 has one genuinely sharp edge: the session SQLite migration is a one-way door for downgrades.
• Sessions created after the migration exist only in SQLite and will not appear if you roll back to an older file-backed release.
• There is a restore path for archived legacy transcripts, but it is a manual step you run from the new CLI before downgrading.

openclaw doctor --session-sqlite restore

Sources: OpenClaw update guide

How to upgrade safely

If you are running OpenClaw in production, here is the path. Read the update guide and the release notes first; the update guide covers the SQLite downgrade trap, the verified-backup step, and rollback paths for every install type.

• Back up first, and verify it. Before upgrading, create a verified backup of your state. Restoring a snapshot is your rollback if anything goes sideways.
• Treat a restore as time travel: approvals and dedupe state roll back too, and messaging credentials, WhatsApp especially, may need relinking.
• Test the upgrade path itself. openclaw update is the recommended path: it detects your install type, runs openclaw doctor, and restarts the gateway.
• On a supervised install, do not update manually in parallel with the running gateway. Stop it first.
• Plan the cutover. If your agents handle live customer conversations you need a rollback plan, and for 2.0 that plan has to account for the SQLite migration, so snapshot everything before you touch it.
• Monitor after the upgrade. The first 24 hours are when you find the bugs testing missed: watch your logs, check your external integrations, and make sure the agent is not silently failing.

openclaw backup create --verify
openclaw update

Sources: OpenClaw update guide · OpenClaw 2.0 release notes

The managed hosting angle

Here is the part we are biased about: if you are running OpenClaw in production, you should not be managing the infrastructure yourself. Qoren handles the upgrades for you. Managed environments run the tested current runtime by default, so when a release like 2.0 drops, upgrades land without you reading migration guides, planning cutover windows, or watching logs at 3am. That is the difference between running agents as a hobby and running them as infrastructure.

• Plans start at $39/mo (Starter: one always-on agent), then $69 (Pro: two agents).
• $169 (Ultimate: five agents, built for agencies running one team per client).
• $499 (Business: ten agents, unlimited scaling, API access).
• Every plan includes BYOK or managed model keys with hard spend caps, all models including frontier, and a 7-day free trial.
• 2.0 is not a tier feature. Every plan runs the current runtime, which is the point of managed.

Should you upgrade?

If you are running OpenClaw 24/7: yes, but carefully, with a verified backup first and the SQLite downgrade trap in mind. If you are just testing: wait a week, let the dust settle, then upgrade. If you are thinking about running agents in production but have not started, now is a good time. The tooling is better than it has ever been, and if 2.0's multiplayer sessions are a preview of where OpenClaw is going, agents are becoming team infrastructure rather than toys. Just do not run it yourself unless you enjoy debugging at 3am. There are people who do that for a living.