Hey :wave: I'm experiencing an issue with an AI a...
# koog-agentic-framework
a
Hey 👋 I'm experiencing an issue with an AI agent using reAct strategy and Gemini Flash 2.5. The issue is that even though the pre-last response in the
onLLMCallCompleted
event appears to not be empty, the agent returns an empty response itself. When I run the same prompt and code with GPT 4.1, the results are always returned correctly. I'd appreciate it if anyone has encountered a similar issue.
v
a
oh, looks like I have to wait with using Gemini then. Google isn't making moves for fixing this. Thanks for your answer! Vadim, in your talks I've never seen Gemini in examples. Do you use it anywhere in JetBrains or does this issue prevent it?
v
I think we provide Gemini in AI assistant — so you can actually set IntelliJ to use it. As for the AI features and other products, I can’t really answer because of, well, NDA 🥲 The issue is only affecting pro2.5 and probably also flash2.5. I would personally use the following logic for myself : GPT models are usually better for conversation, Sonnet is typically better with coding tasks and tool calling, Google’s Gemini models have the longest context (analyzing large documents) which is their main UVP and differentiator.
So I would imagine a situation where you would use GPT4o for chatting, then pass to Gemini (but not 2.5 — maybe 2.0 for example) to read and summarize large documents, then switch to Sonnet to do tool calling
✍️ 1
a
It makes sense, thanks. I guess with Jetbrains AI it works well because all LLM messages are shown in the chat, right? And for my agent it is important to have the last message as the result of the whole agent's work. Following your advice, I'll try Gemini 2.0 Flash for the last step. Thanks a lot 🫶
kodee loving 1