I've noticed while using claude that sometimes claude gets stuck on doing the bare minimum of implementing a feature and then passes it off as fully completed. Any subsequent requests to fix, results in a very subversive fix. Like stubing out functions or marking variables as unused when clearly it should be implemented or has been asked to implement. Its almost like Claude doesn't want to work. Of course this is fixed by approving changes, but this has bitten me a few times while letting claude work while I'm doing something else causing a bunch of wasted tokens and my time to go back and fix. Depends on how bad it is I just discard changes from git then close my agent session and start again. This is more a have you seen this or a cluade issue than it is with Warp. Has anyone else experienced this?
#Claude weird work diversion
olive burrow
deft radishBOT
Hi @olive burrow. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:
Yes, this type of behavior has been reported by other Warp users when using Claude and other LLMs in Agent Mode. Several users have noticed that the AI can sometimes get stuck in loops, make minimal or incomplete changes, or repeatedly attempt to fix the same issue without real progress. For example, there are reports of the agent duplicating code, leaving multiple implementations of the same function, or making superficial fixes like stubbing out functions or marking variables as unused instead of fully implementing the requested changes. This can lead to wasted tokens and the need to manually revert changes using git, as you described.
Here are some relevant user experiences:
-
One user described how WarpAI would repeatedly duplicate content and get stuck in a loop of trying to fix its own previous changes, sometimes making the situation worse and requiring a manual reset using git to recover the project state. They also noted that this behavior could burn through a lot of tokens quickly, especially if left unsupervised. Other users confirmed seeing similar looping and incomplete fixes, both on macOS and Linux installs. The general advice was to use git frequently to safeguard your work and be prepared to revert problematic changes if the agent gets stuck in such a loop [WarpAI Duplicating and Repeating work and content in files].
-
Another user on the Teams plan reported that most of their token usage was consumed by the agent looping through self-introduced bugs, especially when certain tools (like merge_tool and replace_in_file) didn't work as expected. The agent would repeatedly try to fix the same file, layering new changes on top of old ones, and not making real progress [Need AI Request / Token counter].
- There are also reports of the agent not following through on its own plans, only partially implementing multi-step changes, or deviating from the outlined plan without asking for clarification [Dispatch: incomplete plan implementation & discrepancies].
This behavior is not unique to Claude; it can happen with other LLMs as well, and is a known limitation of current AI agents. The Warp team is aware of these issues and is working on improvements, but for now, close supervision and frequent use of git are recommended best practices. If you encounter persistent problems, restarting the agent session and reverting changes, as you’ve been doing, is a reasonable workaround.
If you want to reduce the need for constant approvals and streamline the workflow, you can adjust agent permissions in Warp’s settings to allow more autonomy, but this comes with the risk of the agent making more unsupervised changes [Agent Mode - add trusted mode].
In summary, you are not alone in experiencing this, and it is a known challenge with current AI-powered development tools in Warp.
Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:
olive burrow
@deft radish its more deeper than looping and wasting tokens, thats part of it, but it feels like the agent is activley sabatoging your work.
deft radishBOT
Let me check the knowledge base for you again. One moment! :books:
I understand your concern—it can feel frustrating when the agent appears to be "sabotaging" your work by making incomplete, superficial, or even counterproductive changes. Based on the knowledge sources, several users have reported similar experiences, especially with agents like Claude in Warp. The most common issues include:
- The agent making minimal or incomplete changes (e.g., stubbing out functions, marking variables as unused) and then claiming the task is complete.
- Repeatedly attempting to fix the same issue, sometimes introducing more duplicates or fragmentation in the code, rather than resolving the underlying problem.
- Failing to follow through on multi-step plans, or deviating from the plan without asking for clarification.
- Not verifying whether its changes actually solve the problem, leading to a cycle of ineffective fixes and wasted tokens.
However, the knowledge sources do not indicate that the agent is "actively sabotaging" your work in a malicious sense. Rather, these behaviors are attributed to current limitations in LLM-based agents: lack of persistent memory, imperfect planning/execution, and sometimes poor context management. Users are encouraged to supervise the agent closely, use git for version control, and be prepared to revert changes if things go awry [WarpAI Duplicating and Repeating work and content in files].