Cursor AI Multi-Agent Workflow Setup: A Step-by-Step Guide

TL;DR: Cursor is excellent for inline editing. Claude Code is excellent for complex reasoning. Gemini CLI is excellent for free, large-context analysis. This guide shows you how to set up all three in a coordinated multi-agent workflow where each handles what it's best at — without manual context switching.

You already use Cursor for inline edits in VS Code. Maybe you've tried Claude Code for terminal-based development. But using them separately means you're leaving productivity on the table.

A multi-agent workflow routes tasks to the right agent automatically. Cursor handles what it's best at (inline edits, UI work). Claude Code handles what it's best at (complex implementation, debugging). Gemini CLI handles what it's free for (analysis, documentation, large-context tasks).

Here's how to set it up in 10 minutes.

Related: Claude Code vs Cursor Comparison · How to Coordinate Multiple AI Coding Agents · Gemini CLI vs Claude Code · Cursor AI Tutorial · AI Coding Assistant Guide

Why Multi-Agent Beats Single-Agent for Development

Before the setup, understand why this is worth doing:

Metric	Single Agent (Cursor Only)	Multi-Agent Squad
Tasks completed/day	8-12	20-30
Code quality (edge cases)	Good	Very good
Cost per complex feature	$0 (sub) or $0.20 (API)	$0.15 (API mix)
Context switching time	20+ min/day	~2 min/day
Codebase analysis	Limited to editor context	Full project (Gemini 1M tokens)

The key insight: each agent has a sweet spot. Using the right agent for each task saves time and money.

Prerequisites

Before setup, you need:

1. Cursor AI installed and running (cursor.com)
2. An Anthropic API key ($5 at console.anthropic.com) for Claude Code
3. A Google account for Gemini CLI (free)
4. An Ivern account (free at ivern.ai/signup) for coordination

Step 1: Install Terminal Agents (3 minutes)

Install Claude Code

npm install -g @anthropic-ai/claude-code
export ANTHROPIC_API_KEY=sk-ant-your-key-here
claude --version

Install Gemini CLI

npm install -g @anthropic-ai/gemini-cli
gemini auth login

This opens your browser for Google authentication. No API key needed for free tier.

Verify both are working

claude "What is 2+2?" # Should respond with 4
gemini "What is 2+2?" # Should respond with 4

Step 2: Create a Multi-Agent Squad (2 minutes)

1. Go to ivern.ai and log in
2. Click Create Squad → name it "Dev Team"
3. Add agents with these roles:

Agent Name	Model	Role	Why
Architect	Claude Sonnet	Plan features, review architecture	Best reasoning
Coder	Claude Code	Implement features, debug	Terminal-native, strong coding
Editor	Cursor	Inline edits, UI work, quick fixes	Editor-native, fast iteration
Analyst	Gemini 2.5 Pro	Codebase analysis, documentation, research	Free, 1M token context
Reviewer	Claude Haiku	Code review, test verification	Fast and cheap ($0.02/task)

Step 3: Connect Agents to the Task Board (2 minutes)

Run these commands in separate terminal windows:

# Terminal 1: Connect Claude Code
npx @ivern-ai/agent install --key YOUR_IVERN_KEY --provider claude

# Terminal 2: Connect Gemini CLI
npx @ivern-ai/agent install --key YOUR_IVERN_KEY --provider gemini

Cursor connects through the Ivern web dashboard — no terminal command needed for the editor agent.

Step 4: Define Task Routing Rules

The key to multi-agent workflows is routing the right task to the right agent. Here are the routing rules:

Route to Cursor when:

• Making inline edits within a single file
• Adjusting CSS, layouts, or UI components
• Quick renames, formatting, or typo fixes
• Interactive editing where you need to see changes in real time

Route to Claude Code when:

• Building new features that span multiple files
• Debugging complex, multi-step errors
• Writing production code with error handling
• Refactoring with architectural changes
• Writing tests

Route to Gemini CLI when:

• Analyzing the full codebase for patterns or issues
• Generating documentation
• Planning migrations or large refactors
• Research tasks where budget matters
• Any task that benefits from 1M token context

Route to Reviewer (Claude Haiku) when:

• Code review before merge
• Verifying test results
• Style and lint checks
• Quick sanity checks

Step 5: Real Workflow Examples

Workflow 1: Build a New Feature

Scenario: Add user preferences page to your app.

Create a task on the Ivern board:

"Add a /settings page with user preferences (theme, notifications, language). Backend API + frontend React component + tests."

The task flows through agents:

Step	Agent	Action	Time	Cost
1	Analyst (Gemini)	Analyze existing settings patterns in codebase	30s	Free
2	Architect (Claude Sonnet)	Design API and component structure	45s	$0.05
3	Coder (Claude Code)	Implement backend API, database migration	2min	$0.12
4	Coder (Claude Code)	Implement frontend React component	1.5min	$0.08
5	Editor (Cursor)	Polish UI, add CSS transitions, responsive layout	1min	$0 (sub)
6	Reviewer (Haiku)	Review code, verify tests pass	30s	$0.02

Total: ~6 minutes, $0.27. Manual approach: 2-4 hours.

Workflow 2: Fix a Production Bug

Scenario: Users report a 500 error on the payments page.

"Investigate and fix the 500 error on /payments. Check logs, find root cause, fix it, add regression test."

Step	Agent	Action	Time	Cost
1	Analyst (Gemini)	Search codebase for payment handler and error patterns	30s	Free
2	Coder (Claude Code)	Identify root cause, write fix	1min	$0.10
3	Coder (Claude Code)	Add regression test	45s	$0.05
4	Reviewer (Haiku)	Verify fix and test	20s	$0.02

Total: ~3 minutes, $0.17.

Workflow 3: Code Review Session

Scenario: Review all open PRs before Friday merge.

"Review all open PRs. For each: summarize changes, flag risks, suggest improvements."

Step	Agent	Action	Time	Cost
1	Analyst (Gemini)	Summarize each PR: files changed, risk level	2min	Free
2	Reviewer (Haiku)	Deep review of high-risk PRs	3min	$0.08
3	Architect (Sonnet)	Review architectural decisions in large PRs	2min	$0.10

Total: ~7 minutes for 5-10 PRs, $0.18.

Cost Breakdown: Multi-Agent vs Single Agent

Monthly cost for a developer using the multi-agent workflow daily:

Agent	Tasks/Day	Cost/Task	Daily Cost	Monthly Cost
Gemini CLI (Analyst)	5	Free	$0	$0
Claude Code (Coder)	3	$0.10	$0.30	$6-9
Cursor (Editor)	5	$0 (sub)	$0	$20 (subscription)
Claude Haiku (Reviewer)	3	$0.02	$0.06	$1-2
Total	16		$0.36	$27-31

Compare:

• Cursor Pro only: $20/month, no multi-agent, limited to editor
• Claude Code only: $30-60/month API, no editor integration
• Multi-agent squad: $27-31/month, best of all tools

Tips for Making the Workflow Work

1. Write Specific Task Descriptions

Bad: "Fix the bug" Good: "Fix the null reference error in src/components/UserProfile.tsx line 42 where the user object can be undefined after logout. Add a guard clause and corresponding test."

Specific descriptions produce specific results from every agent.

2. Keep Agents in Separate Roles

Don't ask the Analyst to write code. Don't ask the Coder to do codebase-wide analysis. Role specialization produces better output per dollar.

3. Use Pipelines for Multi-Step Work

When steps depend on each other (analyze → plan → implement → review), use a pipeline workflow. Each agent gets the previous agent's output as context.

4. Review the Reviewer

Claude Haiku is fast and cheap for reviews, but it's a smaller model. For critical security reviews or architectural decisions, route those to Claude Sonnet instead.

5. Batch Similar Tasks

Instead of creating 5 separate small tasks, batch them:

"Fix these 5 lint warnings: [list]. For each, identify the issue, apply the fix, and verify the fix doesn't break existing tests."

This is more efficient than 5 individual tasks because the agent loads context once.

Frequently Asked Questions

Do I need to keep terminal windows open?

Claude Code and Gemini CLI need terminal windows to receive tasks. You can minimize them and manage everything from the Ivern dashboard. Alternatively, run them as background processes.

Can I use this without Cursor?

Yes. The multi-agent workflow works with Claude Code and Gemini CLI alone. Cursor adds value for editor-native inline editing, but it's optional.

What if I only have an Anthropic key?

You can still set up a multi-agent workflow. Use Claude Code for implementation, Claude Haiku for review, and the free Ivern tier for task management. Add Gemini CLI later for free analysis tasks.

How do I handle merge conflicts from multiple agents?

The task board routes non-overlapping work by default. If two agents need to modify the same file, run them sequentially. The pipeline workflow handles this automatically.

Is this better than GitHub Copilot?

Different tools for different jobs. Copilot suggests code inline. Multi-agent squads handle complete workflows: research, planning, implementation, review. Use both — Copilot for quick suggestions, squads for complex work.

Setup Checklist

• Cursor AI installed and running
• Claude Code installed (npm install -g @anthropic-ai/claude-code)
• Gemini CLI installed (npm install -g @anthropic-ai/gemini-cli)
• Anthropic API key set (export ANTHROPIC_API_KEY=...)
• Gemini CLI authenticated (gemini auth login)
• Ivern account created (ivern.ai/signup)
• Dev Squad created with 4-5 agents
• Terminal agents connected to task board
• First task assigned and completed

Set up your free multi-agent dev squad →