Granite 8B passed all 13 parts, yet the car could not drive 1 lap

The article comes from a student who spent the spring of 2026 on an exchange year in the UK, taking an IBM client course that required teams to build a TORCS racing AI while using IBM Granite, IBM's LLM. The class setup put Granite in VS Code through the Continue.dev extension, with Ollama serving the model locally rather than through a cloud API, and the rules also required finishing IBM's online course and showing how Granite was used in a video and blog. The student's team decided in late March to split work by function, have Granite write each part, and have at least two people review any generated code before it was merged in.

What makes the story useful is the gap between mechanical checking and real performance. The workflow moved from manual prompting to a 435-line automated loop that called Ollama's API directly, with the model split between granite-code:8b for code and granite3.2:8b for rewriting prompts. Thirteen parts were defined and checked by per-part Python scripts, with retries, prompt rewrites after two failures, and a Plan B that split stubborn parts further after five failures. The parts passed, but the assembled car did not, because the tests only confirmed fidelity to the specifications, and the specifications contained the actual mistakes.

The article also draws a distinction between being required to use Granite and being allowed to use IBM Bob, IBM's coding agent that became generally available on April 28, 2026. Bob was not used for the racing AI itself but for repository cleanup, and the student describes it as capable enough to plan across multiple files and finish four of five tasks in five instructions, though he still preferred Claude Code. The broader point is about degrees of freedom: a fixed 8B model forces a team to first discover what the tool can do, while an agent that can route between several models is a different kind of companion. The outcome, and the value of the lesson, hinges on whether teams pair part-level tests with a whole-goal test from the outset.