What is OpenClaw? Inside AI Agents, LLMs and the Agentic Loop

Technological Concepts / Core Idea

Gap in typical chatbots

• Standard LLM chat can tell you what to do (e.g., how to schedule a meeting).
• It usually can’t perform actions autonomously in real systems (e.g., editing a calendar or sending emails) without additional infrastructure.

AI agent vs chatbot

An AI agent typically combines:

• an LLM (reasoning),
• tool use (acting),
• and autonomous execution via an agentic loop, rather than a single prompt/response interaction.

Agentic loop / ReAct pattern (core framework)

1. Task arrives to the agent (e.g., from Slack/iMessage/etc.).
2. The agent builds context for the LLM, including:
   • conversation history
   • long-term memory
   • system instructions
   • available tools
3. The LLM reasons and decides whether it needs a tool call.
4. The agent executes tools (e.g., run a command, read a file, call an API, search the web).
5. Tool results are fed back into the LLM’s context window.
6. The loop repeats until the task is complete.
7. The final response is sent back to the user in the original channel.

---

Product / Architecture Details: OpenClaw

What OpenClaw is

• A free open-source AI agent (described as created in late 2025).
• Claimed to be one of the most popular GitHub projects by stars.

Where it runs

• A local Node.js service running on the user’s machine (laptop/VM/Raspberry Pi).

Hub-and-spoke architecture centered on a “Gateway”

• The Gateway is an always-on WebSocket server acting as a control plane.
• It handles:
  • message routing
  • session management
  • creating multiple agents
  • tool usage coordination

Communication channels & adapters

• Integrations include Slack, Teams, Discord, iMessage, and others.
• Adapters convert incoming message formats into a unified internal format for the gateway.

LLM inputs and supporting artifacts

The gateway provides the LLM with more than the user query, including:

• additional context/memory from external storage (long-term memory)
• prompt templates
• markdown instruction files such as:
  • agents.md
  • skills.md

These are described as controlling how the agent should respond and what it should do.

Tools layer

Built-in automation tools include:

• web browser automation
• terminal / CLI execution

Skills layer (extensibility mechanism)

• Skills are extensible units (folders containing markdown skill instructions).
• Key design choice: OpenClaw does not load all skills into the LLM context by default to avoid context-window bloat.
• Instead, it provides:
  • available skills with brief metadata
  • and lets the LLM pull full skill details on demand

Examples mentioned:

• Trello board updates
• Google Calendar access/editing
• Docker workflows (build/run/test container images)
• Integrations with CRMs and GitHub/data sources
• Scheduling automations like cron jobs

---

Security / Enterprise Readiness Considerations (Explicitly Covered)

Local power = local risk

• Because OpenClaw runs locally and can access the file system and terminal, misconfiguration can create a “backdoor” risk.
• The video claims there are many exposed/malicious instances due to:
  • misconfiguration
  • malicious skills code

Prompt injection vulnerability

• Risk occurs when untrusted inputs (e.g., emails/webpages) include malicious instructions the agent may follow.

Mitigations recommended

• Run agents in isolated environments
• Review skills/code
• Encrypt credentials before sending them to an LLM

---

Reviews / Guides / Tutorials

• No specific product “review” score is provided.
• The video is positioned as a technical guide/explainer covering:
  • how agentic loops work
  • OpenClaw’s architecture (gateway, adapters, tools, skills)
  • a security checklist (prompt injection, isolation, credential handling)

---

Main Speakers / Sources

• Single presenter / narrator.
• No names, credentials, or external source citations are provided in the subtitles.