I want to use a UI Agent to enter a simple website, use credentials, navigate the page, and ultimately extract data. The agent actually succeeds in doing so, but it takes a long time because of these errors: “Retrying reporting result and getting next action: attempt N.”
While the logs show these errors, the UI agent doesn’t seem to do anything. Then a “next action thought” log appears and the agent performs an action, which is usually correct. I just want to avoid these retrying errors between actions. The image shows what I mean
Hi @Marcos_Piotto,
This is pretty interesting behavior; so ultimately, the agent is eventually providing the requested information but it takes an extended time due to the errors?
Not sure if applicable or not, but have you encountered this issue when making other requests? Have you only tested with this one specific site for instance?
One other question out of curiosity, what’s the format you’re asking the question in?
There could be several reasons for this. The service is designed to reliably ensure your tasks completes and within a reasonable timeframe. The site could be running into issues navigating the site, connecting, etc.
Here’s what I recommend:
Open up a ticket with support to find out why this behavior is happening
From your end, you can do the following. Often agents struggle with the same repeatable process. e.g. Find login button, find nav for x, find specific form entry, etc. You can build an MCP that will provide the agent context about actions it already took and help during the next time you visit the site. For example, when the agent processes a request for a site you ask it to call “save_memory(site, details)” to ensure it saves the details for the next time it access the site. Then when it runs again the agent will call “get_memory(site)”, to help in aiding in instructions. After that it just involves fine tuning your prompts, and evicting memory when stale.
That’s actually a very clever way to persist memories, will try it out, thanks so much! Actually, I’ve been developing other MCPs to work around different limitations of Quick, but I hadn’t thought of this approach.
PD: Is there a place where I can find more tips like this?
