Simon Willison Benchmarks Codex Desktop Sub-Agents in 3D Game Vibe Coding
Simon Willison tested Codex Desktop using GPT-5.6 Sol Ultra in aggressive sub-agent mode to construct a complete 3D web game from a single prompt. While Codex coordinated sub-agents and generated visual assets using gpt-image-2, it failed to fix visual rendering bugs during screenshot self-inspection.
Impact: Medium
Why it matters
Understand the current performance limits and visual inspection failures of multi-agent execution loops in vibe-coding workflows.
TL;DR
- 01Sub-agent orchestration handles multi-file asset and logic generation effectively
- 02Vision-based self-correction in agent loops still misses prominent visual bugs
- 03Inspect full transcript logs to evaluate sub-agent decision paths and API call costs
Key facts
- Execution Time
- 52 minutes
- Primary Model
- GPT-5.6 Sol Ultra
- Asset Model
- gpt-image-2
Multi-Agent Game Generation
Simon Willison published a project breakdown comparing Claude Fable 5 and Codex Desktop running GPT-5.6 Sol Ultra. Posing a complex single-prompt game specification, Codex operated in aggressive sub-agent mode to coordinate repository creation, gameplay logic, and texture generation via gpt-image-2 over a 52-minute session.
Vision Inspection Limits
While the sub-agent architecture successfully assembled multi-character stacking mechanics and 3D museum environments, it exposed limitations in automated visual inspection. Codex took and analyzed application screenshots during iteration but failed to detect enlarged black eye spheres floating over the models.
✓ When to use
- One-shot prototyping of complex interactive applications and visual assets
- Parallelizing code generation, graphics scripting, and boilerplate test setup
✕ When NOT to use
- Relying solely on visual agent inspection for UI/UX validation
- Unattended production code deployments without automated continuous integration passes
What to do today
- Review Simon Willison's public Codex transcript to evaluate sub-agent task prompts
- Implement explicit unit test assertions for visual properties when relying on coding agents
Sources