Gemini 1.5 Pro for Code - Part 01

Gemini 1.5 Pro for Code can ingest a real GitHub-style repository, then generate working multi-agent Python code that interacts with the repo’s structure—first using OpenAI by default, then switching to Gemini models, and finally adding external tools like DuckDuckGo search. The practical takeaway is that code-focused prompting plus repository context can produce end-to-end prototypes (agents, tasks, and tool calls) with relatively little manual wiring, even when some integration details still need human correction.

The workflow starts by uploading the crewAI repository and selecting the most relevant parts: the source code and the documentation markdown files, while skipping tests. Roughly 35,000–37,000 tokens are included, and the prompt asks Gemini to summarize what crewAI does and what it’s built with. Gemini identifies the stack as Python-based and points to Pydantic and LangChain, along with OpenAI as the default LLM integration. It also produces a concrete list of pip packages needed for a Colab setup (including crewAI and LangChain-related dependencies), which can be copied directly into a notebook.

With dependencies installed, Gemini generates a simple two-agent bot: one agent acts as a hotel-chain customer seeking value for money, and the other plays a salesperson selling air conditioners. The code uses crewAI’s Agent/Task/Crew constructs and runs quickly because the prompt context stays around the 40k-token range. The resulting interaction includes a dialogue-like exchange where the salesperson asks for specifics (e.g., number of rooms and floors) and the system produces a final set of options and requirements. The run also reveals small quirks—like gender-neutral phrasing when the prompt specifies a “saleswoman.”

Next comes model switching. Gemini is asked to modify the same crewAI code to use a Gemini model via Google’s generative AI integration. The generated version largely keeps the structure but introduces integration mistakes—such as an incorrect package name and import path—requiring manual adjustment to use langchain_google_genai and ChatGoogleGenerativeAI. After fixing those details, the bot runs successfully on Gemini, and the outputs differ from the OpenAI version.

Finally, Gemini is pushed to include tool use: a search agent gathers recent AI-release information via DuckDuckGo, and a second agent rewrites it from an “AI doomer” perspective. The first attempt fails due to missing required fields (notably backstory), but after adding backstories, the tool-enabled pipeline works. It retrieves items including OpenAI’s Sora and Google-related generative AI and personalization news, then produces sensational rewrite text.

Overall, the results suggest a strong pattern: repository source code alone can be enough for Gemini to infer key classes, inputs, and outputs, enabling code understanding and downstream generation (tests, additional agents, and possibly docs—though documentation quality may require more guidance).