Change the model client

// Before
BaseLLM llm = PaLMLLM.builder().model("chat-bison").build();

// After
BaseLLM llm = GeminiLLM.builder().model("gemini-2.5-flash").build();

agent.setModel(llm);  // that's it — agent code unchanged
Advertisement

Prompt tuning

Gemini responds better to structured instructions. Add explicit sections: 'Goal', 'Constraints', 'Output format'. PaLM was more forgiving; Gemini rewards structure.

Advertisement

Tool calls now work

PaLM required prompt hacking for tool calls. Gemini has native function calling. Delete your prompt-level tool orchestration.