When maintaining a project as large as OpenClaw, there are hundreds of possibilities for what our maintainers can choose to work on.
As we’ve shipped update after update, we’ve realized these changes were causing difficulties for users, a hard problem made harder by the impossibly large permutations of user configs. But when we released 2.0 in early September, the issue of breaking updates became too large to ignore.
Today, maintainer Jason Sy, who has volunteered his time to OpenClaw from its earliest days, shares more about the update problem we’ve been grappling with, why it was so hard to fix, his recent work to tackle this challenge, and what the end game is for making updates fully seamless.
The scale of configurations
A lot of the feedback was the usual, ‘I updated, it didn’t work,’ and it became a talking point for us during our stand-ups.
Before Jason’s recent fix, there were two possible outcomes when updating your OpenClaw: incremental improvement or catastrophic failure, with no in-between. In the latter case, not only would your update not work, but the old version of your Claw would be down, so there was no agent around to help out with the fix.
I love openclaw and I am using it since early days but Every single update ALWAYS break something for me, Openclaw is a piece of art but one more broken update and you will watch me kill the lobster in all of my systems, and I will be very sad user of Hermes.💔cc: @steipete
— Babikir Yagoub (@babikiryagoub) September 15, 2026
But why was this problem so hard in the first place?
It’s near-impossible to test for every configuration scenario, and as a result, easy to dismiss the update problem as one tied to the wide surface area of the product. With the thousands of config options in OpenClaw, the total number of possible permutations of setups was astronomical.
But when you make a massive improvement to the product, only for it to ultimately be unavailable to people because their claw crashed during an update, it’s pretty disappointing. After the universal update problems surrounding the 2.0 launch, it became ever-clearer that we needed to make repairs immediately.
The developers had a safety net users didn’t
We got so used to it, we didn’t really experience the pain points of just having OpenClaw update itself.
Part of the reason OpenClaw maintainers didn’t feel the pain of upgrades is that they are all developers. As such, they all have some form of coding agent sitting on their computer, and they had all fallen into the pattern of just asking Claude or Codex to update their claw for them.
Asking the agent to fix breakage is a totally rational approach for a developer but doesn’t take into account that for many OpenClaw users, their claw is their only agent. So when their agent goes offline, there is nothing left to fix it.
*03:45 this morning* OpenClaw tried an auto-update (2026.9.1 → 2026.9.2). It shut the gateway down for the handoff, then the restart failed:
— PeteFazz (@PeteFazz) September 6, 2026
```ENOENT ... /opt/homebrew/lib/node_modules/openclaw/dist/shared-DFJEouXv.js```
Jason realized that the answer to fixing upgrades lay within the pattern that he and the other maintainers had adopted. Now he had to find a solution.
Atomic updates and easier issue reporting
It should just always leave you with a working claw to at least diagnose what issues happened with the update and potentially fix it on its own. Even if the system goes down, the agent should still keep going and push that update through.
It turned out that all the pieces to make agentic-assisted upgrades happen smoothly already existed within the current upgrade process. They were just in the wrong order.
Our new structure of atomic updates works like this:
- OpenClaw will keep your existing gateway running while it prepares for the update.
- If the update fails, it will roll back to the last working configuration.
- OpenClaw will always preserve a working agent so that it can help you diagnose and potentially fix any failures that happen during upgrades.
Finally! A self update that doesn’t break OC. Good progress @openclaw
— Drouzzart (@Teckkariuki) September 10, 2026
Alongside this change, we’ve also launched an easy issue reporting button for updates.
The most important thing now, if you’re reading this, is to use that button and jump in our Discord if you’re ever experiencing a problem. We are committed to making this process better so that everybody can get the best experience on the most recent versions of OpenClaw.
Feedback, plus what’s still ahead
I don’t wanna say that updating is completely fixed or anything, but from what we can gather, it is working in a lot of scenarios. If it’s not, jump in our Discord, report an issue, ping me. That type of feedback is definitely very useful.
In the future, the goal is for updates to be so easy and seamless that users don’t even notice them. What this may look like is an agent-assisted sidecar that’s detached from your agent itself and exists solely to monitor your upgrade process and fix any changes, report any issues, and handle the entire process for you.
When you have an open-source product with such a vast range of setup scenarios, it’s hard to say that upgrades will always be perfect, or fixed. Still, since launching atomic updates, we’re starting to hear more and more positive signs from the community. Likewise, we’ve already received feedback from the issue reporting that we’ve acted on to resolve.
First impression of OpenClaw v9.3 is great. Got my Discord tool calls back and seems snappier than 8.2.
— Mike Assad (@mikeassad77) September 10, 2026
It was also the first time an agent led update worked flawlessly for me.
If you’ve been waiting for a solid “OpenClaw 2.0” release, this is the one.
We’ve made significant progress on our updates process and are excited for people to try. If you haven’t upgraded your OpenClaw recently because you’ve been burned in the past, we totally understand. But there’s never been a better time to be running the latest version of OpenClaw.
We hope you’ll give it a look. And if you have an issue, there’s now a button for that.