Output variable of Chat block no longer working as of 10pm May 5 2026

[michael_r_blum]

This breaks every agent I have been working on and every workflow.
Please get back to me as soon as possible.

When using the Chat block the variable used to capture and persist data during the workflow is broken.

This started happening at 10pm.

Every new agent I create is impacted, every new workspace I create is impacted, every computer I try is impacted, both PC and Mac.

BUT:
I have an agent that must be using an old version of the sandbox because it works fine.

STEPS TO REPRODUCE:

After the bug was found I created a new test agent with a ‘Chat’ block.
I create a variable named ‘chat_history’ that should save the chat to a variable.

When I try to reference it in the ‘Generate Text’ block intellisense is now missing showing the sandbox has lost reference to it:

Debugger shows bug:
‘chat_history’ is being saved in the block but never can be referenced after:

I made a copy of an agent published yesterday before the 10pm breaking update.
It still works, ‘chat_history’ can be referenced in the workflow:

I created a new workspace, a new user with a new workspace and tried a different computer with each of those scenarios.
All broken.
Something was pushed last night that now impacts the entire platform.

NOTE: test agent was created by Architect V2 as part of bug troubleshooting and bug persists, meaning Architect can no longer create workable agents.

The chat block and its output is the first link in every workflow I am building and with this broken my entire system is broken.

This is URGENT URGENT URGENT!!!

I will make myself available to ZOOM or Google Chat if you would like to screenshare, I can demo the bug.

Please help.

Link to broken Agent (draft workspace):

https://app.mindstudio.ai/agents/c4fe06d5-d289-4a14-a67c-c0d71cfb18a3/edit

Link to working agent:

https://app.mindstudio.ai/agents/c7118a20-340d-4a07-8e17-3b7b62d43ba0/edit

[Alex_MindStudio]

Hi Michael, thanks for the post!

Could you publish the broken agent in your workspace and enable remixing? That’ll let me create a copy in my workspace to test.

Also, could you share the Debugger logs for that failed run? Here’s how:

1. Open the Debugger
2. Select the run
3. Click Share
4. Reply with the URL
   
   

[michael_r_blum]

remixing has been enabled.

Failed run link:

https://app.mindstudio.ai/share/debugger/c4fe06d5-d289-4a14-a67c-c0d71cfb18a3/dfd6e7ec-8ba5-408f-a0a1-5448572f7d7f

NOTE:

The failed run link opened a new tab that does not show the state I see.
My failed run debugger shows the red squiggle error with hover text:

[Alex_MindStudio]

Hi @michael_r_blum,

Thanks for sharing those!

Looking at the debugger log, the run doesn’t show any messages sent during that chat session, which would explain why chat_history comes back as undefined.

Could you run it again, send a message in the chat, and then share the new debugger link? That’ll confirm whether the block is actually failing to capture, or whether the test run just didn’t have any input to work with.

It would also help if you could record the steps you’re taking:

Click here to record your screen

[michael_r_blum]

Thank you Alex for being so helpful.
I did submit a video - it ran long but I wanted to give you every detail I could think of.

[Alex_MindStudio]

Hi @michael_r_blum,

Thanks for the recording! The mic was muted, so if you were narrating, I’m afraid it didn’t come through.

Here’s what I can see: the user types a message and hits Submit Request, which transitions the workflow to the next block right away. The Chat block only logs messages to the history variable when the user clicks Send, which actually starts the chat. Since no message was sent before moving on, there are no logs to capture.

If you’re looking to collect a single message, I’d recommend using a User Input block instead.

[michael_r_blum]

apologies for the no audio - I dont have a mic on my PC so I’m forced to use the macbook. I enabled the mic in privacy settings for chrome but the recorder you linked me to refuses to use the mic.

Back to the bug:

And that is the thing.. I know you know your system and believe me I am so excited and impressed by it.
But yesterday I could simply type a command and that would trigger a workflow by clicking ‘Submit Request’.

This is what was supposed to be narrated:
We want to connect small businesses to AI - delegating menial, repetitive tasks.

The agent you saw is private to the client.

When the client wants to start a workflow at the click of a button they type in the trigger and hit Submit Request.

This hits the Dispatcher which detects the users intent and generates a JSON payload that first hits the SQLite DB to grab any data needed and then is routed to the correct ‘Run Workflow’ block - in this case the social media poster.

The social media poster will be one of many public facing generic agents that rely on JSON payload to inject all dependencies needed to run.

The part that hasn’t been built yet is what happens after the social media poster returns with its output.

I had made it this far I swear I only had to hit submit request without chatting first.

I wouldn’t have built the other links in the chain if I Submit Request is behaving as it is now.

Submit Request is the first link and now the whole idea is in jeopardy.

If you want to look at it here is the url.
I enabled remixing but what are the rules there? Do I have to publish?

https://app.mindstudio.ai/agents/31118bcd-ddaa-4e3c-96c8-9e9aa567ae39/edit

What I also forgot to mention:

I tried to move the purple ‘Set Variable’ blocks as global variables because I am still a bit new and didn’t know I had global vars available to me.

The buggy behavior started after that.

I was unable to reference the global vars with a {{global.var}}. Display output would show the string literal not the variable value.

So I deleted the global vars and put the Set Variable blocks back.

Immediately after is when the bug started.

I thought of using the ‘User Input’ block but that ruins the user experience.

The agents has a system prompt with a harness with User, Soul, Style etc.
The client is supposed to be able to chat if they want to and execute workflows with a separate button.

[michael_r_blum]

Hi Alex - As I was typing out the last edit our interaction gave me an idea.
I added a ‘Trigger Phrase’ section to the system prompt:

So the client can chat and then execute a trigger from there.

Dispatcher block shows the correct decision made:

So we found a work around. Thank you.

Apologies for the verboseness I hope you can understand from developer to developer how high strung we get.

[Alex_MindStudio]

Hi @michael_r_blum, thanks for sharing the additional details.

Yes, the agent needs to be published before remixing works.

I’ve checked with our engineers and can confirm there were no recent updates to the Chat block. The behavior around history variable logging has been this way for some time, and messages only get logged once sent in chat.

From what you’ve described, the Chat block isn’t the right fit for this workflow. Your users aren’t having a back-and-forth conversation with an LLM, they’re typing a trigger and moving on. A User Input block would be a much better fit for your use case. It captures the input, passes it forward, and doesn’t require a chat session at all.

I’d also like to flag that with the Chat block, the user always has to click Next to end the chat session before the workflow continues to the next blocks, regardless of what’s in the System Prompt.

[michael_r_blum]

Thank you for checking - maybe I am the one hallucinating now..

Surreal, I could have sworn… but more importantly you are right.

I like the flow of the user being able to chat with a ‘Submit Request’ button next to it.
When the user wants to start a workflow they click it which sends them to the ‘User Input’ block and the user asks for the workflow there.
Conceptually much easier.
And the LLM interaction won’t taint the explicit request.

Tests are working correctly now.

I was just married to that first idea but this one is better.

Thank you for your patience Alex - drop a link if I can leave a review as someone who found Mindstudio responsive and helpful to their community.

[Alex_MindStudio]

Hi @michael_r_blum,

That flow is still possible with the Chat block! The Submit Request button will always be visible, but it should only be clicked after the user has sent at least one message in the chat. If someone types a message without sending it and clicks Submit Request right away, nothing will be logged.

As for the variables, global variables persist across runs for all users, so you may not need Set Variable blocks at all. We have a quick tutorial on the different variable types here:

Types of Variables

Depending on your use case, Dynamic Tools might be a good fit. If you want users to stay in the chat interface and let the LLM decide when to trigger a specific action, that’s exactly what it’s designed for. Here’s a walkthrough of the full setup:

MindStudio Update: Dynamic Tool Use in Chat Block | Sean Thielen posted on...

And here’s a sample agent you can remix to explore the structure:
https://app.mindstudio.ai/agents/dynamic-tools-in-chat-8650c49b/remix

[michael_r_blum]

wow thank you @Alex_MindStudio for all the info.

I was unaware of the Dynamic Tools - I am excited to dive into it and see how they apply to what I am building.