Building Maze Atelier with Gemini 3.8 Flash and Boost

The Breakthrough

Google Antigravity introduced the /boost workflow, which allows users to trigger multi-agent reasoning for complex coding tasks by delegating sub-tasks like implementation and investigation to specialized subagents.

What Actually Worked

• Workflow Integration: The author initiated the project by selecting the Gemini 3.8 Flash model and prefixing the prompt with /boost to trigger the multi-agent strategy layer.
• Constraint-Based Prompting: To ensure game logic integrity, the author defined specific failure cases for the agent, including blocked movement, missing keys, unreachable paths, and state management during solver resets.
• Subagent Verification: The system identified and resolved bugs related to event subscription drops during maze regeneration and movement validation errors that previously allowed coordinate jumps through walls.
• Regression Testing: The final build included a suite of seven automated tests covering collision detection, key requirements, blocked grid states, and a batch run of 100 generated mazes to ensure consistency.

Context

The author built "Maze Atelier," a procedural maze game using plain HTML, CSS, and JavaScript with canvas, to test the limits of Gemini 3.8 Flash and the /boost workflow. The project required handling complex state transitions, such as updating pathfinding when a user manually edits walls. While the author successfully generated a functional game, they noted that /boost did not appear in the autocomplete during the test, and they emphasized that usage costs depend on individual plan quotas and the volume of agent work performed rather than a fixed multiplier.