AI Agents Overview - Automagik Suite Documentation

Overview

Forge supports 8 different AI coding agents - giving you the freedom to choose the best tool for each task. This is the core of Forge’s BYOL (Bring Your Own LLM) philosophy.

The Agent Landscape

Supported Agents

Commercial API-Based

Claude Code

By: Anthropic Models: Claude 3.5 Sonnet, Opus, Haiku Best for: Complex logic, architecture Cost: $3-15/MTokExcellent reasoning, great for system design

Cursor CLI

By: Cursor Models: Multiple (via Cursor account) Best for: UI/UX, rapid iteration Cost: Subscription-basedIntuitive for frontend work, fast iterations

Gemini

By: Google Models: Gemini 2.0 Flash, 1.5 Pro Best for: Fast tasks, experimentation Cost: Free tier available!Blazing fast, generous free tier

OpenAI Codex

By: OpenAI Models: GPT-4 Turbo, GPT-3.5 Best for: General purpose coding Cost: $0.50-30/MTokReliable, well-documented

Open Source & Local

OpenCode

Type: Open source Run: Locally via Ollama Best for: Privacy-sensitive work Cost: Free (your hardware)Fully local, no data leaves your machine

Qwen Code

Type: Open source (Alibaba) Run: Locally via Ollama Best for: Cost-conscious development Cost: Free (your hardware)Strong performance, competitive with commercial

LLM-Agnostic

Claude Code Router

Type: LLM proxy Use: Any OpenAI-compatible API Best for: Flexibility, vendor independence Cost: Varies by backendRoute to ANY model - OpenAI, Anthropic, local, custom

Amp

By: Sourcegraph Type: Code intelligence Best for: Large codebase navigation Cost: Based on usageExcels at understanding existing code

Choosing the Right Agent

By Task Type

Task Type	Best Agent	Why
Complex Architecture	Claude Sonnet/Opus	Superior reasoning
Quick Fixes	Gemini Flash	Fastest, free
UI Components	Cursor	UI/UX intuition
Refactoring	Claude Sonnet	Maintains coherence
Testing	Claude/GPT-4	Comprehensive coverage
Documentation	GPT-4	Clear writing
Privacy-Critical	OpenCode/Qwen	Runs locally
Experimentation	Gemini	Free tier, fast

By Budget

💰 Budget-Conscious:

Gemini Flash (free tier!)
OpenCode (local, free)
GPT-3.5 Turbo (cheap)

💵 Balanced:

Claude Haiku (fast + cheap)
Gemini Pro
Cursor (subscription)

💎 Premium:

Claude Opus (best quality)
GPT-4 Turbo
Claude Sonnet

By Context Window

Agent	Context	Best For
Gemini 1.5 Pro	2M tokens	Massive codebases
Claude 3.5	200K tokens	Large projects
GPT-4 Turbo	128K tokens	Standard projects
GPT-3.5	16K tokens	Small files

Agent Comparison

Speed vs Quality

Quality
  ↑
  │     Claude Opus ●
  │            Claude Sonnet ●
  │                    GPT-4 ●
  │                         Cursor ●
  │                              Claude Haiku ●
  │                                   Gemini Pro ●
  │                                        GPT-3.5 ●
  │                                             Gemini Flash ●
  └──────────────────────────────────────────────────────────→
                                                          Speed

Cost vs Capability

Capability
  ↑
  │     Claude Opus ●
  │            GPT-4 ●
  │                Claude Sonnet ●
  │                     Gemini Pro ●
  │                          Claude Haiku ●
  │                               GPT-3.5 ●
  │                                    Gemini Flash ●
  │                                         OpenCode/Qwen ●
  └──────────────────────────────────────────────────────────→
                                                          Cost
                                                      (cheaper)

Real-World Performance

Benchmark: “Add JWT Authentication”

Same task, different agents:

╭───────────┬──────────┬─────────┬────────┬───────╮
│ Agent     │ Duration │ Quality │ Tests  │ Cost  │
├───────────┼──────────┼─────────┼────────┼───────┤
│ Claude S  │ 5m 22s   │ A+      │ 95%    │ $0.23 │
│ Gemini F  │ 2m 15s   │ B+      │ 78%    │ $0.00 │
│ GPT-4     │ 6m 01s   │ A       │ 88%    │ $0.45 │
│ Cursor    │ 4m 18s   │ A-      │ 85%    │ $0.19 │
│ OpenCode  │ 8m 33s   │ B       │ 70%    │ $0.00 │
╰───────────┴──────────┴─────────┴────────┴───────╯

Winner: Claude Sonnet (best balance)
Budget: Gemini Flash (free!)
Speed: Gemini Flash (2m 15s)

Switching Between Agents

One of Forge’s superpowers: try multiple agents on the same task!

# Start with fast, cheap agent
forge task create "Add authentication" --llm gemini

# If not satisfied, try Claude
forge task fork 1 --llm claude

# Compare results
forge task compare 1

# Choose winner
forge task merge 1 --attempt 2

Agent Configuration

Quick Setup

See detailed setup for each agent:

Configuration File

.forge/config.json:

{
  "llms": {
    "claude": {
      "apiKey": "sk-ant-...",
      "model": "claude-3-5-sonnet-20241022"
    },
    "gemini": {
      "apiKey": "AIza...",
      "model": "gemini-2.0-flash-exp"
    },
    "openai": {
      "apiKey": "sk-...",
      "model": "gpt-4-turbo"
    },
    "cursor": {
      "enabled": true
    }
  }
}

Specialized Agent Profiles

Apply “personas” to any base agent:

# Security-focused Claude
forge task create "Add auth" \
  --llm claude \
  --agent "security-expert"

# Test-focused Gemini
forge task create "Add feature" \
  --llm gemini \
  --agent "test-writer"

# Performance-focused GPT-4
forge task create "Optimize query" \
  --llm openai \
  --agent "performance-optimizer"

See Specialized Agents for details.

Agent Strengths & Weaknesses

Claude Code

Strengths ✅:

Complex reasoning
System design
Edge case handling
Clear explanations

Weaknesses ⚠️:

Can be verbose
Sometimes over-engineers
More expensive

Best for: Architecture, refactoring, security

Gemini

Strengths ✅:

Blazing fast
Free tier (!)
Good for simple tasks
Concise code

Weaknesses ⚠️:

May miss edge cases
Less sophisticated reasoning
Shorter responses

Best for: Quick fixes, iteration, experimentation

Cursor CLI

Strengths ✅:

Great UI/UX intuition
Fast iterations
Context-aware
Good for frontend

Weaknesses ⚠️:

Subscription required
Less depth on algorithms
Tied to Cursor ecosystem

Best for: UI components, rapid prototyping

OpenAI Codex (GPT-4)

Strengths ✅:

Reliable and consistent
Well-documented
Good all-rounder
Strong community

Weaknesses ⚠️:

Expensive
Slower than alternatives
Nothing exceptional

Best for: General coding, documentation

Open Source (OpenCode/Qwen)

Strengths ✅:

Fully local
Privacy guaranteed
Free (your hardware)
No rate limits

Weaknesses ⚠️:

Slower
Lower quality
Requires powerful hardware
More setup needed

Best for: Privacy-sensitive, learning, cost control

Multi-Agent Workflows

Sequential Workflow

Use different agents for different stages:

# 1. Design with Claude (best at architecture)
forge task create "Build payment system" --llm claude

# 2. Implement with Cursor (fast iteration)
forge task create "Build UI components" --llm cursor

# 3. Test with GPT-4 (comprehensive tests)
forge task create "Add integration tests" --llm openai

# 4. Document with Gemini (fast, cheap)
forge task create "Write API docs" --llm gemini

Parallel Comparison

Run multiple agents simultaneously:

# Try 3 agents at once
forge task create "Optimize database query" --llm claude &
forge task fork 1 --llm gemini &
forge task fork 1 --llm openai &

wait

# Compare and choose best
forge task compare 1

Cost Tracking

Monitor spending per agent:

# This month's costs
forge cost summary --month january

# Output:
╭──────────┬─────────┬────────╮
│ Agent    │ Tasks   │ Cost   │
├──────────┼─────────┼────────┤
│ Claude   │ 24      │ $5.67  │
│ Gemini   │ 48      │ $0.00  │
│ GPT-4    │ 12      │ $8.23  │
│ Cursor   │ 15      │ (sub)  │
╰──────────┴─────────┴────────╯
Total: $13.90

Best Practices

Start with Free Tier

Begin with Gemini Flash (free!) to validate approach, then use premium agents for refinement

Match Agent to Task

Don’t use Claude Opus for simple fixes. Don’t use Gemini Flash for architecture.

Learn Agent Strengths

Track which agent works best for which task types in your codebase

Keep Options Open

Configure multiple agents. Vendor lock-in is the enemy of productivity.

Troubleshooting

Agent not available

Error: “Agent ‘claude’ not configured”Solution:

Check .forge/config.json has API key
Verify API key is valid
Run forge config validate

All agents slow

Issue: Even “fast” agents are slowPossible causes:

Network latency
Large context window
Complex task description

Solutions:

Check internet speed
Reduce context
Simplify task description

Results inconsistent

Issue: Same task, different results each timeThis is normal!AI is non-deterministic. Use temperature=0 for more consistency:

{
  "llms": {
    "claude": {
      "temperature": 0
    }
  }
}

Next Steps

Claude Code Setup

Configure Claude for Forge

Gemini Setup

Setup Gemini (free tier!)

Open Source Agents

Run agents locally

Specialized Agents

Create custom agent personas

​Overview

​The Agent Landscape

​Supported Agents

​Commercial API-Based

Claude Code

Cursor CLI

Gemini

OpenAI Codex

​Open Source & Local

OpenCode

Qwen Code

​LLM-Agnostic

Claude Code Router

Amp

​Choosing the Right Agent

​By Task Type

​By Budget

​By Context Window

​Agent Comparison

​Speed vs Quality

​Cost vs Capability

​Real-World Performance

​Benchmark: “Add JWT Authentication”

​Switching Between Agents

​Agent Configuration

​Quick Setup

​Configuration File

​Specialized Agent Profiles

​Agent Strengths & Weaknesses

​Claude Code

​Gemini

​Cursor CLI

​OpenAI Codex (GPT-4)

​Open Source (OpenCode/Qwen)

​Multi-Agent Workflows

​Sequential Workflow

​Parallel Comparison

​Cost Tracking

​Best Practices

Start with Free Tier

Match Agent to Task

Learn Agent Strengths

Keep Options Open

​Troubleshooting

​Next Steps

Claude Code Setup

Gemini Setup

Open Source Agents

Specialized Agents

Overview

The Agent Landscape

Supported Agents

Commercial API-Based

Open Source & Local

LLM-Agnostic

Choosing the Right Agent

By Task Type

By Budget

By Context Window

Agent Comparison

Speed vs Quality

Cost vs Capability

Real-World Performance

Benchmark: “Add JWT Authentication”

Switching Between Agents

Agent Configuration

Quick Setup

Configuration File

Specialized Agent Profiles

Agent Strengths & Weaknesses

Claude Code

Gemini

Cursor CLI

OpenAI Codex (GPT-4)

Open Source (OpenCode/Qwen)

Multi-Agent Workflows

Sequential Workflow

Parallel Comparison

Cost Tracking

Best Practices

Troubleshooting

Next Steps