v20: Agent looping/repeating actions in agentic workflows
Hi! First thank you for all the hard work. Really appreciate it!
Testing v20 with cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4 via vllm v0.20. v20 is a big improvement over v19 (no longer needs constant nudging), but I'm hitting a persistent looping issue during agentic workflows (Claude Code / Hermes).
The Problem: Agent Spins in Circles
The agent repeatedly dispatches the same background agents or repeats the exact same steps multiple times before finally agreeing it's done. Sometimes during these loops it breaks the code it was trying to fix.
Example 1 — Re-dispatching the same agent over and over:
⏺ Agent "Commit 1: LLM repair cascade" completed · 2m 24s
Thought for 6s
⏺ The agent was denied tool permissions. Let me implement both commits directly.
... [does edits] ...
⏺ The previous agent was denied permissions. Let me re-dispatch both commits.
⏺ Agent(Commit 1: LLM repair cascade) — Backgrounded
...
⏺ Agent "Commit 1: LLM repair cascade" completed · 2m 23s
Thought for 59s
⏺ I'll dispatch agent 1 for Commit 1 now, then follow up with agent 2.
⏺ Agent(Commit 1: LLM repair cascade) — Backgrounded
...
⏺ Agent "Commit 1: LLM repair cascade" completed · 2m 19s
Thought for 55s
⏺ The previous agent was denied tool permissions, so I'll dispatch agent 1 for Commit 1 now.
⏺ Agent(Commit 1: LLM repair cascade) — Backgrounded
... [repeats 4–5 times total]
The agent "forgets" it already dispatched and completed the task, then immediately re-dispatches it. This loop runs 4–5+ times per task.
Example 2 — Repeating the same direct edits:
⏺ The agent was denied permissions. Let me implement both commits directly.
[edits file]
⏺ The agent was denied permissions. Let me implement both commits directly.
[edits same file again]
⏺ I see - the agent was denied permissions. Let me implement both commits directly.
[edits same file again]
It repeats the same edit block multiple times, sometimes corrupting the file in the process.
My Setup:
- Model:
cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4 - Engine: vllm v0.20
- Kwargs:
preserve_thinking: true(thinking enabled, traces kept in context)
Has anyone else seen this spinning behavior with v20? Any tips to mitigate this?
Thanks!
I actually had the same problem yesterday in OpenCode. The session was not that long and after a few edits this happened. It completely broke the session and I was not able to continue, because it started outputting malformed calls or looping.
Same issue, not on agent tasks though, but it basically seemed to be calling every tool, deleted some stuff that shouldn't have been deleted.. I didn't seem to recall having this issue on v19 unfortunately:/
I can confirm, when you put it that way, that the model feels significantly dumber than before. I "didn't notice" and was thinking that it is just the code, that is more complex or simply more of it. But when I think about it more, the code si relatively simple, size is similar. I had multiple issues for the past few days, where it removed important parts of a process and left it for me to debug why is the code behaving differently now. I am at a point where I could be writing the code myself actually and it would probably be faster and less frustrating. 😀 I have downgraded to v19 for now and I will be observing the behavior and let you know as well.
Hi! The massive v21 update was just released. It completely overhauled tool-calling compatibility (switching to native Hermes JSON for inference engines like llama.cpp and LM Studio), fixed the preserve_thinking amnesia stalls/loops, and resolved several </think> parsing and prompt injection bugs.
This issue should now be fully resolved in v21. I am closing this discussion, but please feel free to reopen or create a new one if you are still experiencing any trouble. Thanks!