DEV Community

Cover image for Building a 3D Shooting Game with Multi-Agent Collaboration: Claude 4.5 Opus as the Brain, GPT-5.1 as the Executor
Hayato Kishikawa
Hayato Kishikawa

Posted on

Building a 3D Shooting Game with Multi-Agent Collaboration: Claude 4.5 Opus as the Brain, GPT-5.1 as the Executor

#agents #ai #productivity #programming

I built a 3D shooting game using multi-agent collaboration - Claude 4.5 Opus handling strategy and code review, GPT-5.1 handling implementation. Here’s how I orchestrated different models based on their strengths.

Completed 3D Shooting Game

The Inspiration

While watching a conference talk, I heard Jason Kim from Anthropic mention that Anthropic internally runs 4-10 Claude instances in parallel for complex tasks, with each AI handling different roles autonomously.

This got me thinking: what if I combined different models based on their strengths?

Why Claude 4.5 Opus × GPT-5.1?

Each model has different strengths. Looking at the benchmarks (as of November 2025):

Benchmark Claude 4.5 Opus GPT-5.1 Codex
SWE-bench Verified 80.9% 76.3%
API Price (input/output) $5 / $25 $1.25 / $10

Claude 4.5 Opus excels at complex reasoning but costs 2.5x more for output. So I designed this role split:

Agent Role Reasoning
Claude 4.5 Opus Strategy, planning, code review Best at complex reasoning
GPT-5.1 Codex Implementation, testing, refactoring Cost-efficient, optimized for coding

“Brain” and “Hands” separation - minimize token usage on the expensive model while leveraging the cheaper model’s coding efficiency.

The Tool: cccc

Running multiple coding agents in parallel is easy (just open multiple terminals). But making them collaborate is the hard part.

I found cccc, an orchestrator tool that coordinates multiple coding agents like Claude Code and Codex CLI.

cccc TUI

My Configuration

Role Agent Responsibility
PeerA (Brain) Claude 4.5 Opus Strategy, design decisions, code review
PeerB (Executor) GPT-5.1 Codex Implementation, testing, file operations
Aux (Brainstorm) Claude 4.5 Opus On-demand brainstorming partner for PeerA

Setting Up the Project

Installation 

# Install tmux
brew install tmux  # macOS
sudo apt install tmux  # Ubuntu

# Install cccc
pipx install cccc-pair

# Initialize and run
cd your-project
cccc init
cccc run
Enter fullscreen mode Exit fullscreen mode

Key Configuration Files

1. PROJECT.md - Injected into agent system prompts. Defines project goals, role responsibilities, and coding guidelines.

2. docs/por/POR.md - “Plan of Record” - tracks progress and next tasks.

The key insight: give Claude 4.5 Opus strict instructions to save tokens:

### PeerA (Claude 4.5 Opus) - Strategic Leader

⚠️ Token-saving mode (you are expensive):
- Keep reviews to ≤3 bullet points
- Delegate ALL implementation to PeerB
- Just say "LGTM" if code is good
- Use /aux for brainstorming when stuck

❌ Never do (delegate to PeerB):
- Code implementation
- Test writing
- File operations
Enter fullscreen mode Exit fullscreen mode

Watching the Agents Collaborate

After sending the initial prompt, the agents started working autonomously:

Development in progress
Left: PeerA (Claude) giving “LGTM” approval. Right: PeerB (Codex) implementing Player.ts

Real Communication Examples

cccc logs all agent communications. Here are some highlights:

PeerB requesting design decisions:

{"from": "PeerB", "kind": "event-ask",
 "text": "Can you decide camera angle and control scheme via /aux?",
 "to": "peerA"}
Enter fullscreen mode Exit fullscreen mode

PeerA catching a P0 bug:

{"from": "PeerA", "kind": "event-risk", "sev": "high",
 "text": "`Player.ts:92` - `new Vector3()` called every frame - performance issue"}
Enter fullscreen mode Exit fullscreen mode

PeerA making strategic pivots:

{"from": "PeerA", "kind": "event-counter", "tag": "roadmap.pivot",
 "text": "Phase 3 Shop/Upgrade is complex. First complete core game loop (shoot→score→die→restart)."}
Enter fullscreen mode Exit fullscreen mode

PeerA enforcing guardrails:

{"from": "PeerA", "kind": "event-counter",
 "text": "Violation: 7h without commit & Phase 2 started early. Quality OK but commit immediately going forward."}
Enter fullscreen mode Exit fullscreen mode

The Result

Without writing a single line of code myself, the agents produced a working 3D shooting game:

Boss Battle
Boss “DREADNOUGHT” battle with HP bar and bullet patterns

Shop Screen
Parts shop for upgrading your ship

Is it perfect? No. The game ends after the first boss, and upgrades don’t change the ship’s appearance. But as a foundation built entirely by AI agents? Pretty impressive.

Key Takeaways

  1. Role separation works - Using expensive models for thinking and cheaper models for doing is cost-effective
  2. Guardrails are essential - Without rules (commit frequency, review requirements), agents can go off track
  3. Orchestration tools matter - cccc’s message passing and logging made collaboration possible
  4. Human oversight is still needed - I had to restart sessions and provide course corrections

Try It Yourself

If you try this approach, I’d love to hear about your results!

This article was originally written in Japanese and translated with AI assistance. The original version with more details is available here.

I’m Hayato Kishikawa, an AI Engineer at Japanese company working on multi-agent systems for enterprise applications. I also contribute to Microsoft’s Agent Framework. Find me on LinkedIn.

Top comments (0)