Multi-Agent Setup: Reduce Rate Limits with Antigravity + Cerebras

[chindris-mihai-alexandru]

Overview

Following the suggestion from @NoeFabris on PR #258, I'm sharing this multi-agent configuration pattern here for community visibility.

Problem: Antigravity rate limits can be hit frequently during heavy development sessions.

Solution: Use Antigravity only for complex reasoning tasks, and delegate fast/simple tasks to free providers like Cerebras.

---

Architecture

Primary Agent (Antigravity Claude Opus)
    ├── explore (Cerebras - fast, free)
    ├── researcher (Cerebras - fast, free)
    ├── coder (Cerebras - fast, free)
    └── security-auditor (Antigravity Sonnet - needs thinking)

Model Tiers

Tier	Provider	Models	Use Case
Heavy	Antigravity	antigravity-claude-opus-4-5-thinking, antigravity-claude-sonnet-4-5-thinking	Complex reasoning, architecture
Fast	Cerebras	llama-3.3-70b, qwen-3-32b	Exploration, research, code generation

---

Example Configurations

Primary Agent (Antigravity)

<!-- ~/.config/opencode/agent/antigravity.md -->
---
model: google/antigravity-claude-opus-4-5-thinking
mode: primary
maxSteps: 100
---
You are an elite coding agent with access to Claude Opus 4.5 Thinking.

Fast Subagent (Cerebras)

<!-- ~/.config/opencode/agent/explore.md -->
---
model: cerebras/llama-3.3-70b
mode: subagent
tools:
  - Read
  - Glob
  - Grep
---
Fast agent for codebase exploration. Read-only access.

Security Auditor (Antigravity - needs thinking)

<!-- ~/.config/opencode/agent/security-auditor.md -->
---
model: google/antigravity-claude-sonnet-4-5-thinking
mode: subagent
tools:
  - Read
  - Glob
  - Grep
---
Security analysis agent. Thorough review of code for vulnerabilities.

---

Why This Works

• Primary agent uses Antigravity for complex multi-step reasoning
• Fast subagents use Cerebras (free, ~2100 tokens/sec) for quick tasks
• Rate limits hit less often since most requests go to free providers
• Quality maintained - complex work still uses Claude thinking

---

Related

• OpenCode Agent Docs
• oh-my-opencode #647 - Configure default variants from config
• anomalyco/opencode #6645 - Support for Model Variants in subagents

---

Would love to hear if others have tried similar setups or have improvements to suggest!