#Destructive agent behavior

10 messages · Page 1 of 1 (latest)

velvet whale
#

I have experienced destructive behavior from agents, during working sessions involving code modifications. Imagine you have worked for some hours and already got some good progress. Then, the agent misunderstands some of your instructions or simply, all of a sudden, gives up and decides to backtrack and delete some files or changes. Your progress is (partially/completely) lost. At the same time, you wasted your AI request/usage. Angry/frustrated customer.

What can be done to be able to time travel along your work history, and avoid loosing valuable progress/ideas?

frosty sphinxBOT
#

Hi @velvet whale. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:

#

I'm sorry to hear about your experience with destructive agent behavior in Warp. Based on the knowledge sources, here are the key points and recommendations to help you avoid losing valuable progress and to "time travel" along your work history:

1. Use Version Control (e.g., Git)

Warp team members and users strongly recommend working within directories that are under version control (like Git). This allows you to revert changes, recover deleted files, and track your progress over time. If an agent deletes or modifies files unexpectedly, you can restore them using your version control history. This is the most robust way to safeguard your work against accidental loss by AI agents or any other process No Human in the Middle for File Deletions!!!.

2. Configure Agent Mode Autonomy and Denylists

You can control agent autonomy in Warp:

  • Go to Settings > AI > Autonomy.
  • Ensure destructive commands (like rm) are in the command denylist. By default, Warp includes risky commands in this list, so they always require your confirmation before execution.
  • You can also disable model-based auto-execution, so all commands require your approval before running.

This setup helps prevent agents from executing destructive actions without your explicit consent No Human in the Middle for File Deletions!!!.

3. Manual Intervention

If you notice the agent attempting a destructive action, you can use CTRL-C to cancel the operation immediately.

4. Current Limitations

At this time, the knowledge sources do not mention a built-in "time travel" or undo feature within Warp itself (such as session snapshots or checkpoints). The best available method is to rely on external version control systems for this purpose.

#

5. Feedback and Improvements

The Warp team is aware of these concerns and is actively working on improving Agent Mode's behavior, including the possibility of adding checkpoints or additional safeguards in the future No Human in the Middle for File Deletions!!!.

Summary:
To avoid losing progress, always work in a version-controlled directory, configure agent autonomy and denylists, and be ready to intervene manually. There is currently no built-in Warp feature for session "time travel," so version control remains your best tool for recovering lost work.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:

#

@urban spoke Kapa.ai is still learning and improving, please let me know how I did by reacting below.

velvet whale
#

I use git all the time. But agent could probably do automatic commits of important/big changes because sometimes you're not able to stop it...

little maple
#

You can issue it an instruction to make sure it commits after every change.

#

I try to use the remember command. ie remember to commit changes after each major change to the code.

#

The remember command helps with the Context Window limits too. Let it remember your important things regardless of the context window letting stuff fall out of the conversation.

velvet whale
#

Rules! Yes, I need a good set of rules to get better results during my working sessions with agents.