Bug: AI Agent Randomly Switching Identity
AI Agent randomly ignores prompt/KB and switches to completely different identities (new names, jobs, companies).
Report:
• Issue: AI agents are randomly assuming different identities during conversations.
• Expected Behavior: Agent should consistently follow the system prompt and knowledge base (example: Name = Ana, Company = XYZ Technology).
• Actual Behavior: At random points, the agent introduces itself with unrelated names and roles (e.g., “Liz from a dental clinic,” “Sofia from an AI agency”).
• Frequency: Happens unpredictably, not tied to a specific input.
• Impact: Creates inconsistent and confusing user experience.
• Question: Is this a known issue, and how can I ensure the agent always respects the defined identity?
Attached there are 2 screenshots from different agents where the problem happened today (Sep 13):
- In the one, where we see the name Liz, should be Ana. URGENT - because our customers is already using it.
- In the other, where we see Sofia, should be Laura. It is still in development.
6 Replies
which LLM is the agent using?
The agent is using Claude 3.5 - Haiku.
In another I assigned to the KB the Claude 3.5 Haiku, and in the workflow it is using Claude 4 Sonnet.
for my own agent steps, i find that Claude 3.5 - Haiku still sometimes doesn't quite follow the prompt/KB, so i needed to upgrade to Claude 4 Sonnet
I'll follow your recommendation. Thank you, @chirashi !!
just to clarify, while changing to Claude 4 definitely can help the agent step follow the prompt/KB better than Claude 3.5, i am not necessarily saying that this will fix your entire problem about identity changes
Ok. Thanks @chirashi for making clearer your recommendation! Let's see if it will fix the problem. Otherwise, I'll update you too.