How Omnigent Unified Scheduling Gained 4,000+ Stars in 5 Days for Multi‑Agent Coding

Problem

Developers often use multiple AI coding agents (Claude Code for business logic, Codex for API lookup, Cursor for debugging, Gemini for research). Each agent runs in an isolated session, causing context fragmentation, duplicated prompts, token waste, tool isolation, uncontrolled costs, difficult collaboration, and security risks such as accidental exposure of secrets.

Core Concepts

Harness is the wrapper around a specific agent (Claude Code, Codex, Cursor, Pi, OpenAI Agents). It encapsulates prompts, file operations and tool calls but does not interoperate with other harnesses.

Meta‑Harness (Omnigent) is a unified scheduling layer built on top of all harnesses. It acts like a vehicle chassis where individual harnesses are interchangeable engines.

Architecture

Omnigent consists of two modules:

Runner – Provides an isolated sandbox for each sub‑agent. Linux uses bubblewrap, macOS uses the system seatbelt. Each sub‑agent runs in its own Git worktree, guaranteeing independent execution and a standardized I/O interface.

Server – Manages global policies, session storage, multi‑user collaboration, cross‑device synchronization and exposes terminals, WebUI, mobile and API endpoints so that a single session is reachable from all devices.

Key Capabilities

Polly – Code Generation & Review

Task decomposition – automatically splits a requirement into modular sub‑tasks.

Parallel dispatch – assigns sub‑tasks to different vendor agents in isolated sandboxes.

Cross‑review – Claude‑generated code is reviewed by Codex, and GPT‑generated code is validated by Claude.

Human merge – all diffs are aggregated for manual confirmation before merging.

Debby – Dual‑Model Research Assistant

Side‑by‑side comparison – displays two model answers in parallel columns.

Debate mode – the /debate command triggers multi‑turn critique to converge on a single solution.

Custom Agent via YAML

Agents can be defined with a concise YAML file, optionally generated by AI. Example:

name: 数据分析Agent
prompt: 专业数据分析师，输出标准化统计报表
executor:
  harness: claude-sdk  # 切换 codex/cursor 仅改这一行
tools:
  word_count:
    type: function
    callable: mymodule.word_count
researcher:
  type: agent
  prompt: 检索行业数据并精简总结
  tools:
    word_count: inherit

Omnigent can auto‑generate the full YAML from a natural‑language request.

Cross‑Device Session Sync

Sessions synchronize in real time across terminals, WebUI and mobile devices, enabling a developer to start a task on a workstation, continue debugging in a browser after work, and monitor progress on a phone during a commute.

Three‑Layer Policy & Security

Cost control – Soft budget alerts and hard limits pause sessions when spending caps are reached; token usage is tracked per session.

Permission control – File whitelist restricts agents to the src directory; secret files such as .env are blocked.

Action interception – Operations like npm install require manual approval before git push.

OS‑level sandbox – Keys are injected via outbound proxies; agents cannot read credentials directly.

PII protection – Outbound requests are scanned for phone numbers, email addresses, ID numbers and blocked for manual review.

Team Collaboration Modes

Share view – Teammates watch the full agent execution and comment on code or files in real time.

Co‑drive – Teammates join the session and issue commands that run on the host machine, supporting pair programming and remote debugging.

Fork – Duplicate a session to run alternative solutions in parallel without interference.

Engine Switching

Changing the underlying model only requires editing the harness field in the YAML; the rest of the workflow remains unchanged.

Installation Guide

Prerequisites:

Python 3.12+

Node.js 22 LTS + npm

git, tmux

Linux: bubblewrap (macOS has built‑in sandbox)

Method 1 – One‑Click Automated Install (recommended)

curl -fsSL https://raw.githubusercontent.com/omnigent-ai/omnigent/main/scripts/install_oss.sh | sh

Method 2 – Manual Channel Install

# uv tool install omnigent
uv tool install omnigent
# pip install
pip install omnigent
# Homebrew (macOS)
brew install omnigent-ai/tap/omnigent
# Source install
uv tool install -q --python 3.12 git+https://github.com/omnigent-ai/omnigent.git

Basic commands:

# General start and model key setup
omni
omni setup
# Launch specific agents
omni claude
omni codex
# Run built‑in examples
omni run examples/polly/
omni run examples/debby/
# Start server / WebUI
omni server start
# Register host for remote access
omni host
# Fork a session
omni run --fork <session_id>
# Attach to a shared session
omni attach <session_id>
# Upgrade
omni upgrade

Model keys can be configured interactively with omni setup or via environment variables ( ANTHROPIC_API_KEY, OPENAI_API_KEY). Example for local Ollama (Codex engine): BaseURL http://localhost:11434/v1, key is any string.

Known Issues (Alpha)

Occasional pipeline blockage – sub‑agents may error without real‑time alerts; time‑outs up to 2 minutes; reduce parallelism for complex tasks.

High resource usage – 500 MB–1 GB RAM per multi‑agent job; recommend ≥16 GB RAM.

Limited third‑party agent support – only Claude Code, Codex, Cursor, Pi are officially supported; custom adapters required for others.

Learning curve – basic usage is straightforward, but advanced custom policies and multi‑stage workflows require thorough reading of the documentation.

Linux sandbox dependency – missing bwrap causes startup failure; install via apt install bubblewrap.

Network restrictions – GitHub script downloads and model gateways may timeout in regions with restricted access; use proxy mirrors.

Comparison with Single‑Agent Setup

Multi‑model collaboration : Single agents require manual copy‑paste and suffer context breaks; Omnigent schedules parallel sub‑agents in one session and automates cross‑review.

Cost control : Single agents have no real‑time budgeting; Omnigent provides three‑layer budgeting with soft alerts and hard limits.

Permission security : Prompt‑based constraints are easy to bypass; Omnigent enforces OS‑level sandboxing and stateful policies.

Cross‑device sync : Single agents run locally only; Omnigent syncs terminal, web and mobile in real time and can be cloud‑hosted.

Team collaboration : Single agents only forward results; Omnigent offers share view, co‑drive and fork modes.

Migration cost : Switching tools requires rewriting prompts and tool configs; Omnigent switches engines by changing one YAML line ( harness).

Privacy protection : Single agents send PII in clear to models; Omnigent scans outbound traffic and blocks sensitive data.

Open Source Links

https://github.com/omnigent-ai/omnigent

https://omnigent.ai/