Here’s my honest feedback after actively using Cursor with Agent Mode on a real Golang project.
Agent Doesn’t Respect Context
Even though I have a Golang backend project, when I ask the agent to write compatible code, it keeps generating Flutter code inside the Go project—despite repeatedly clarifying that I only need the Flutter code to integrate externally, not within the same structure.
Weak Code Memory
The agent seems to understand and generate initial code correctly, but when I later ask for a feature addition or modification, it behaves as if it doesn’t remember or fully understand the existing structure.
For example, if I have a working API that stores text, and I ask to add image support, it often breaks the text functionality, focusing only on the image — instead of combining both.
Ignores Instructions to Preserve Working Code
Despite explicitly asking not to touch the working parts, it sometimes modifies them when the conversation gets long. This leads to unnecessary bugs and regressions.
Poor File Awareness
When I tell it to read and update its memory with all files in the current project, it often reads only a few files and misses critical components — which later results in incorrect or incomplete code suggestions.
Repeats Mistakes
If it writes faulty code and I ask it to fix the exact issue, it often repeats the same mistake instead of reconsidering a different approach or acknowledging the previous failure.
Agent Mode is Risky for Live Projects
Based on my experience, Agent Mode is risky for active or large codebases. There’s a real chance of:
- Breaking working code
- Generating redundant or mismatched components
- Not fully understanding existing logic
- Failing to collaborate across frontend/backend boundaries properly
I hope this feedback helps improve Agent Mode — it’s promising, but needs more context awareness, instruction adherence, and reliable memory of existing files to be truly powerful.