Been working on an Agent Builder workflow for ages. It’s deployed. Works fine.
As of today though, clicking on it to edit it in Agent Builder gives me a red error “Failed to load workflow” and the workflow does not appear.
So…. ???
1 Like
Doing some more checking on my own, I figured out how to view an older version via URL params and that does load. So basically, some sort of Agent Builder bug preventing certain workflows from loading but no specific error, etc.
Not sure the best way to report this bug. I tried the support email but it seems very consumer focused, not developer.
I was able to get another workflow to fail to load but not sure which node did it. MCP maybe.
Still not sure how to report this.
I have the same error, can you publish that workflow that cannot load?
how you can view the older version via URL? I can’t even do that
I’m stuck on this for a week now, I’m very appreciate if you can help me @hunter.hillegas
@h6503 I’m still working through this. Support actually did/are trying to help me but they said it was fixed and it is not so still working on that.
What I think is happening is that some edit I made somehow corrupted the data or something - i.e. I triggered some bug in the Agent Builder with a change so that now when I click to open it and it tried to load the most recent draft version, that fails and instead of showing me the version before that or anything else, it just errors out.
In my case, I have a few previously published versions - it appears that just the most recently created auto-draft is broken.
So if you have previously published, you can try this too. If you haven’t, not sure what to do other than try support I guess.
A URL like this will open whatever version is in that field. In my case I wasn’t sure what version was my most recent so I just started with “1” in that field and then it loaded and I could pick from the dropdown. Draft is still broken.
https://platform.openai.com/agent-builder/edit?version=X&workflow=wf_XXXXXX
1 Like
Thank you very much @hunter.hillegas , I have to clear the cache and use incognito tabs to have the workflow back.
I tried to contact support but they haven’t answered too. I have exact same error, draft is broken, and I cannot publish anything new one day.
@h6503 Sounds like we’re seeing the same thing. I have a few of these workflows with this problem - I don’t think I was doing anything too “out there”.
Other than using the URL trick above to access an old version which you can use as a reference to copy, not really…
I’m still working with support but I’d encourage you to open your own issue.
I’m having the same issue. the URL trick did not work either. This is a production build for a big client, it was untouched for a while and working fine, just came back to this issue. So very deeply concerning as I can’t even duplicate it or anything.
1 Like
I will keep updating based on what I learn but strongly suggest everyone email support at openai to start a case (and tell the bot to skip you to a human, bot can’t fix this). If this is all the same issue, we really need them to fix it.
I talked to them through the chatbot and it said it send an email. Haven’t found this super helpful in the past though…
I still had a copy version of my project open on an old tab so i was able to edit it a bit.. I tried recreating the whole thing manually, took ages → still broke
What i did instead, is remove each module 1 by 1.
after hours i realised it is the Yellow MCP modules that are the culprit for me.
Updating the Auth etc fiddling with them does nothing; the simple existence of the yellow MCP modules totally breaks the workflow!
update: using an MCP like make.com worked.
it was an MCP with ‘custom headers’ , more specifically Supabase, that is breaking it for me.
It has the project url in supabase and then custom headers with Authorization, Bearer (PAT token here)
update 2: Updating the supabase mcp to not use custom headers and instead use the ‘api key access token’ connection, still breaks. this time it doesnt give the error ‘failed to load workflow’, it literally just opened a new project called new workflow just as when it usually said ‘failed to load workflow’, just without the error message.
update 3: removing the api token completely from the supabase mcp module causes the workflow to work normally again. however of course now i cannot access the supabase MCP as needed by my workflow. hmm.
can someone tag an openai employee or someone with power to this thread? really critical error for me @OpenAI_Support
@Emanuel_tb Excellent detective work. I also have MCPs with custom headers so that sounds like a common thread.
And I hear you on email support but I have found that if I’m really persistent, I eventually get a human who does try to help (though slowly, I get about one reply per day and in the last one they wanted me to try options in the Agent Builder that I do not see to try to resolve it).
I will append this to my existing support case and see if that helps. And yeah, if someone from OAI that sees this can escalate it or make it known to the team working on this, that would be incredible… because this is a huge drag.
I somehow managed to get it working in the most bizarre method, but I cannot confidently say it will stay working.
Essentially, I still had access to a duplicated version of my flow. I rebuilt this from scratch. After doing that, the MCP errors as above happened.
I made a new PAT token for my Supabase mcp, and pasted it into the API key field. As above, that still broke.
What I did next was bizarre - i set it to ‘no authentication’. For some reason when I did this and published it, it appeared to work AND had my api key filled in via ‘Access Token / API Key’….
So now I just have to connect this new workflow to replace the old broken one that I cannot even open. I am so lucky I had a minimized tab of my workflow open for the last week or so, lol, otherwise I would not have been able to recover my project…
If this breaks again I will update
2 Likes
I’m also experiencing the same issue. Any workflow I create that contains an MCP node with custom headers can not be reloaded once it is closed.
This is makes it impossible to use MCP nodes with custom headers in agent builder. @OpenAI_Support
I’m experiencing exactly the same issue when using a workflow that includes an MCP with custom headers.
This has happened to me twice already with different workflows.
Is there any workaround or temporary fix available for this?
I put a floating MCP node in my draft that I’m not even using right now. 
After a reboot last night all my work is just … gone.
We really need @OpenAI_Support to fix it… so if you know anyone that works there on this stuff, let them know.
