Not being able to see what the agent is doing is making me nervous

Support Bug Reports
1

I used the visible agent thoughts collapsed tab to watch and deep monitor per task in order to catch problems before they spread through my codebase. The agent often starts with a wrong assumptions almost 5 of 10 requests this happens , then chooses the wrong tools, changes the wrong files, and builds more code on top of that mistake

This is not just about wasting a few minutes. The output can look correct on the surface while introducing hidden logic, bad architectural decisions, unexpected dependencies, or behavior that is extremely difficult to notice later. By the time the final result appears, the damage may already be scattered across several files.

Previously, I could see where the agent was heading, stop it early with the most feature i love i stop and send new input mid work, and provide the missing context. Now it feels like I am giving it a blind task and hoping it does not quietly compromise the integrity of the project.

I understand that raw internal reasoning may not always be appropriate to display, but users still need a clear, real-time view of the agent’s assumptions, plan, tool choices, files being changed, and intended next steps. Without that level of visibility and control, it is difficult to trust the agent on any serious codebase.

This change is making me heavily reconsider whether this is still the right IDE for my workflow

i can’t tell you how many hundreds times i do this same pattern and agent on frontier models ends up saying actually this is correct i miss understood this

AND NOW WE CODE BLIND ?

6

Hi @abdoomalik Thank you for the post. I didn’t notice any changes that we made to the reasoning, agents still share their thoughts and you can review them (I just verified, and I frequently am looking to do the same as you). Reasoning hasn’t been taken away. The agent still streams its thinking, tool calls, and edits, and you can still stop it mid-run and send new input. What changed is how densely the transcript groups activity by default in the Agents window, so the detail now sits behind a click.

Two things that should help:

  1. Click the chevron on the collapsed “Thinking” row while the agent is running. It expands in place and streams live.
  2. Go to Settings > Appearance > Agent Conversations > Tool Call Density and drag toward Detailed. Edits and terminal commands then render inline as they occur rather than folding into the group.

Let me know if you have any follow up questions!

7

thanks a lot yes found on settings and made it detailed
This seemingly minor feature can affect your entire codebase and save you days of tracing problems later

8

Awesome, thanks for confirming!