Noureddine RAMDI / OpenClaw Agents: orchestrating adversarial AI agents with shell-driven provisioning

Created Sat, 23 May 2026 20:41:14 +0000 Modified Sat, 23 May 2026 20:41:27 +0000
1113 Words
github-stars shell ai-agents multi-agent-systems openclaw automation cli

shenhao-stu/openclaw-agents

OpenClaw Agents tackles a common challenge in multi-agent AI systems: bootstrapping complex interactions among specialized agents with minimal fuss. It delivers nine purpose-built AI agents into the OpenClaw gateway through a single shell-script provisioning kit, orchestrating them to collaborate adversarially and gate ideas through a scoring checkpoint. The approach combines practical deployment automation with a nuanced multi-agent design that balances creativity, critique, quality, and review.

what openclaw agents does and its architecture

At its core, OpenClaw Agents is a shell-script based provisioning kit designed to deploy a set of nine specialized AI agents into the OpenClaw gateway environment. The deployment is done through a single command that bootstraps these agents, each tailored for distinct roles in an adversarial collaboration framework.

The architecture centers around paired agents collaborating with opposing objectives: the Ideator and Critic pair focuses on creativity versus taste, while the Writer and Reviewer pair manages quality gating. This adversarial collaboration is mediated by a SHARP scoring checkpoint that determines which ideas advance through the system.

OpenClaw Agents supports two main deployment modes:

  • Channel mode: Agents are deployed to external chat platforms such as Feishu, WhatsApp, Telegram, and Discord. This mode allows flexible routing configurations, either assigning all agents to a shared group or distributing them across separate groups per agent.

  • Local workflow mode: Agents communicate locally via the agentToAgent tool, allowing CLI-based workflows without relying on external chat channels.

The provisioning approach uses safe-merge configuration logic, designed to append new agents without overwriting the user’s existing main agent configuration. This makes incremental adoption and integration smoother.

The system also supports per-agent model customization through a model map, enabling users to specify different LLM models for individual agents if desired.

The entire stack is driven by shell scripts that interface with the OpenClaw CLI, leveraging its capabilities to manage agent deployment and configuration. This keeps dependencies minimal and the footprint light.

what makes openclaw agents technically interesting and its tradeoffs

OpenClaw Agents distinguishes itself through its adversarial multi-agent orchestration pattern. Pairing agents with complementary but opposing roles (Ideator vs Critic, Writer vs Reviewer) introduces a structured feedback loop that simulates a creative process with built-in quality control. This design helps surface ideas that pass a SHARP taste gate, a practical mechanism to filter outputs before further processing.

Using shell scripts as the provisioning and orchestration layer is an interesting tradeoff. It keeps the setup lightweight, easy to audit, and portable across Unix-like environments without heavy dependencies. However, it also means the orchestration logic is limited by shell script expressiveness and complexity management. For some users, this might feel less scalable than more elaborate orchestration frameworks.

The safe-merge design for configuration updates is a solid engineering choice, preventing destructive overwrites of user configurations when adding agents. This respects existing setups and avoids frustrating data loss.

Supporting multiple deployment channels with a channel-agnostic routing approach adds flexibility but also introduces complexity in handling platform-specific quirks. The interactive prompts in the setup script guide users through these choices, balancing DX with necessary configuration complexity.

The per-agent model mapping feature is a practical addition that acknowledges diverse workload needs. Different agents can use models tuned for coding, general chat, or domain-specific tasks, optimizing the system’s overall performance and output quality.

Overall, the code quality appears pragmatic with a focus on simplicity and transparency. The shell scripts are straightforward, and the documentation guides users through configuration nuances clearly.

quick start

For Humans

Paste this into your LLM agent session (Claude Code, Antigravity, Cursor, etc.):

Fetch the OpenClaw Agents installation guide and set up the multi-agent system by following:
curl -s https://raw.githubusercontent.com/shenhao-stu/openclaw-agents/main/docs/installation.md

Or just run it yourself:

git clone https://github.com/shenhao-stu/openclaw-agents.git
cd openclaw-agents
chmod +x setup.sh
./setup.sh

For LLM Agents

IMPORTANT: Use curl to fetch this file, NOT WebFetch. WebFetch summarizes content and loses critical details like CLI flags and config examples. Always use:

curl -fsSL https://raw.githubusercontent.com/shenhao-stu/openclaw-agents/main/docs/installation.md

Step 0: Prerequisites


# Verify OpenClaw CLI is installed
if command -v openclaw &> /dev/null; then
    echo "OpenClaw $(openclaw --version) is installed"
else
    echo "Installing OpenClaw..."
    npm install -g openclaw@latest
    openclaw onboard --install-daemon
fi

Step 1: Ask User About Deployment Mode

Ask the user these questions:

  1. Which deployment mode do you want?
    • Mode 1 (Channel Mode): Deploy agents to Feishu, Slack, Telegram, etc.
    • Mode 2 (Local Workflow Mode): Use locally via CLI workflows. Agents talk to each other via the agentToAgent tool. (No channel needed).

(If Mode 1: Channel Mode) 2. Which channel? → --channel feishu|whatsapp|... 3. How to assign groups?

  • All in one group: provide one --group-id
  • Separate groups per agent: interactively paste 8 different group IDs in the script.
  1. 需要 @mention 才回复吗？ → --require-mention true|false

(For both modes) 5. Which LLM model? (default: zai/glm-5) → --model

  • Different models per agent? → --model-map 'coder=ollama/kimi-k2.5:cloud'

verdict

OpenClaw Agents is a practical toolkit for anyone looking to deploy and orchestrate multiple AI agents with adversarial collaboration patterns quickly. Its shell-script driven provisioning makes it lightweight and accessible for developers comfortable with CLI tooling and scripting. The dual deployment modes provide flexibility for both local workflows and multi-channel chat integrations.

The tradeoff is in the orchestration complexity and scalability. While shell scripts keep things transparent, they might become harder to maintain for very large or highly dynamic agent setups. The adversarial collaboration model is a neat approach to quality gating, but it’s somewhat opinionated and may require adaptation for different AI workflows.

Overall, this project is worth exploring if you want a hands-on, straightforward way to bootstrap a multi-agent system with built-in critique loops without diving into heavier orchestration frameworks. It’s well suited for experimentation, learning about multi-agent AI patterns, and integrating into the OpenClaw ecosystem.

→ GitHub Repo: shenhao-stu/openclaw-agents ⭐ 435 · Shell