Skip to content

povli/AgentGOD-oss

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AgentGOD

English | 中文

Hierarchical multi-agent orchestration for Cursor IDE.

You talk to one commander. It breaks down your request, delegates to specialist agents, and delivers the combined result — all inside Cursor.

License: MIT Cursor Agents


How It Works

You: "Refactor the utils module, research best practices first, then implement, then review"

                         ┌─────────────┐
                         │ Orchestrator │ ← You talk here
                         └──────┬──────┘
                    ┌───────────┼───────────┐
                    ▼           ▼           ▼
              ┌──────────┐ ┌────────┐ ┌──────────┐
              │Researcher│ │ Coder  │ │ Reviewer │
              │(explore) │ │(build) │ │(inspect) │
              └──────────┘ └────────┘ └──────────┘
                    │           │           │
                    └───────────┼───────────┘
                                ▼
                         Combined result → You

Each agent runs in its own fresh context window via Cursor's Task tool. The orchestrator coordinates everything — parallel execution, error handling, context checkpoints, and even asking you questions when agents need clarification.


Quick Start

1. Clone

git clone https://github.com/povli/AgentGOD.git
cd AgentGOD

2. Deploy (installs global components + initializes your project)

./scripts/deploy.sh /path/to/your-project

3. Open the project in Cursor and start talking

"Help me implement a REST API with authentication"

The orchestrator handles the rest.


Features

Hierarchical Orchestration

One commander agent receives your request, analyzes complexity, and delegates to the right specialists. Simple tasks are handled directly; complex tasks are decomposed and dispatched in parallel.

Fresh Context Per Agent

Each specialist runs in a separate context window (via Cursor's Task tool). No context pollution. When a task gets too long, agents automatically checkpoint their progress and continue in a new context.

Agents Ask You Questions

When an agent lacks critical information, it sends a structured NEEDS_INPUT request back to the orchestrator, which presents the questions to you in a friendly format. Your answers are injected into the agent's continuation.

Project Takeover

Drop the agent system into an existing codebase and run a full scan:

./scripts/deploy.sh --takeover /path/to/existing-project

The scanner agent analyzes directory structure, tech stack, git history, and documentation — then generates a project-knowledge.md that gives the orchestrator full project context.

Fully Customizable Agents

Create new agents by copying a template:

cp agents/_template.md agents/my-specialist.md
# Edit the file — done. No restart needed.

Each agent is a single Markdown file with YAML frontmatter:

---
name: my-specialist
role: One-line description
expertise: [area-1, area-2]
tools: [Read, Write, Grep]
subagent_type: generalPurpose
model: fast  # or leave empty to inherit the main window model
can_ask_user: true
---

You are [role], specialized in [domain].

(system prompt: responsibilities, workflow, output format)

Cross-Project Reusable

Global components install once to ~/.cursor/ and work across all projects. Each project gets its own agent team. Deploy to a new project in one command.


Default Agents

Agent Role Best For
researcher Information search & exploration Codebase analysis, tech research, best practices
coder Code implementation Feature development, bug fixes, refactoring
reviewer Code review & quality Quality checks, security audit, best practices
scanner Project takeover Codebase scanning, knowledge extraction
system-editor System reconfiguration Redesign the agent team for any industry or workflow

Architecture

~/.cursor/ (global, shared across all projects)
├── skills/agent-orchestrator/     ← Orchestration logic
│   ├── SKILL.md
│   └── references/
└── rules/agent-orchestrator.mdc   ← Checkpoint protocol

your-project/ (per-project)
├── .cursor/rules/
│   └── agent-system.mdc           ← Activates the orchestrator
├── agents/                         ← Your agent team (customize freely)
│   ├── researcher.md
│   ├── coder.md
│   ├── reviewer.md
│   └── scanner.md
└── workflows/
    ├── project-knowledge.md        ← Generated on takeover
    └── state/                      ← Runtime state (gitignored)

Deploy Commands

Command What It Does
./scripts/deploy.sh Install global components only
./scripts/deploy.sh /path/to/project Global + init project
./scripts/deploy.sh --takeover /path/to/project Deploy + mark for auto-scan
./scripts/deploy.sh --project-only /path/to/project Init project only
./scripts/deploy.sh --force Force-update global components

All commands are idempotent — running twice won't overwrite your customized agents.


Creating Custom Agents

Step 1: Copy the template

cp agents/_template.md agents/api-designer.md

Step 2: Edit the frontmatter

Field Required Description
name Yes Lowercase + hyphens, 3-50 chars
role Yes One-line description
expertise Yes Areas of expertise (orchestrator uses this for task matching)
tools Yes Allowed tools (principle of least privilege)
subagent_type Yes explore / generalPurpose / shell / browser-use
model No fast for cheap tasks, empty to inherit main model
can_ask_user No Allow the agent to ask you questions (default: false)

Step 3: Write the system prompt (the body of the file)

Step 4: Done. The orchestrator discovers new agents automatically on the next conversation.


How Context Switching Works

Agent starts task
    → Reaches a checkpoint (phase boundary / large output / 8+ tool calls)
    → Saves progress to workflows/state/{id}.md
    → Returns CHECKPOINT to orchestrator
    → Orchestrator spawns a NEW Task that reads the state file
    → Agent continues from where it left off

This happens automatically. You just see the orchestrator say "continuing..." — no action needed from you.


FAQ

The orchestrator isn't delegating tasks? Check that .cursor/rules/agent-system.mdc exists and agents/ has at least one non-_ prefixed .md file.

How many agents can run in parallel? Up to 4 (Cursor Task tool limit). Dependent tasks run sequentially.

Do I need to restart after changing agents? No. The orchestrator scans agents/ dynamically each conversation.

Will this affect projects without AgentGOD? No. The global Skill only activates when trigger conditions match. The global Rule only fires when workflows/ files are present. Projects without agent-system.mdc are unaffected.

How to upgrade?

git pull
./scripts/deploy.sh --force  # Updates global components

How to uninstall?

rm -rf ~/.cursor/skills/agent-orchestrator
rm -f ~/.cursor/rules/agent-orchestrator.mdc
# Per project: remove .cursor/rules/agent-system.mdc, agents/, workflows/

Contributing

Contributions are welcome! Feel free to:

  • Add new agent templates
  • Improve the orchestration logic
  • Fix bugs or improve documentation

Please open an issue first to discuss significant changes.


License

MIT

About

Hierarchical multi-agent orchestration for Cursor IDE. One commander, multiple specialists, zero overhead.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages