I dynamically create the agent call within my workflow.
But when you look into the Agent logs, you see that the mustache is not resolved correctly:
Is that a bug? Or am i doing something wrong?
Additionally is there a way to overwrite the instructions. I am using Messages here, but technically i would like to adjust the agends instructions within the invocation.
Is it possible to overwrite them like:
{
”instructions”: “my dynamic instructions”,
"action": "invoke"
}
Hi @Tobias_Sell,
To confirm, the issue that in the Agents log it shows the input as {{code1.data}} ?
And the agent's run is incorrect because it is not able to parse out the data being passed in?
Not sure about overwriting the instructions, would have to check in with @kent on that.
Hi @Tobias_Sell,
Just wanted to check in, are you still experiencing this bug?
It looks like you are passing in the data correctly in the workflow's agent block call.
For the second part of your question, I don't believe Agent instructions can be dynamically updated from workflow block invocations. Agent instructions are meant to be robust and added at the Agent level via the GUI.
We decided to move the workflow to n8n. The agent actions involve a lot of busy waiting for the agent. Since retools billes by agent hours and we have to pay for the waiting, its way cheaper to push the podcast generating job to n8n. But i personally enjoy the retool way to do it more
Hey @Tobias_Sell,
For what is worth, what you're missing in your code 1 block is to stringify what you're sending, so simply add what tou already have to JSON.stringify(), or stringify it directly in your template block (if that is a code block).
I agree that agents can consume a lot of time. What model did you try? I had a situation where changing from gpt-5.2 to 4.1 reduced the time by three times, so worth trying as well.
We currently are very happy with claude-sonnet-4.5 but it doesn’t rally matter, since the retool agent triggers workflows that need time for processing elsewhere (elevenlabs for audio, cutting api, uploading, downloading files). So we spend money for pretty much having the agent sitting arround and waiting for these workflows to complete.
Out of curiosity, why do you need an Agent rather than doing everything with workflow blocks? I get the impression that your flow is deterministic
elevenlabs for audio, cutting api, uploading, downloading files
EDIT: I guess the action may not always be the same and you depend on the agent to decide which actions to take, which is fair enough.
Arguably, you could have a simple AI block which serves as a branch and each branch goes to a very specific workflow for the actions you need and then it recursively goes back to the initial workflow, but that's the whole point of having an agent, right? Anyway, I just cringe at the sound of n8n so I always do everything I can to stay wihtin retool if possible.
