About This Article

🤖 A Robot That Only Talks vs. an AI That Runs Errands

If you ask ChatGPT "What's the weather like in Seoul today?", it will give you an answer. But that's it. It doesn't actually search the internet to find out, set an alarm to remind you to bring an umbrella, or recommend clothes for the weather and open a shopping site.

Agentic AI is different. It handles all of these things on its own, like a friend who's great at running errands. If you say "Help me prepare for tomorrow's presentation," it can search the internet for relevant materials, organize them, and even create a file for you.

🦊 The Story of pb

At Pebblous, there's an AI friend called pb. When you send pb a message on WhatsApp, it responds. But it doesn't just respond with words.

If you say "Organize this week's meeting schedule," pb opens Google Calendar on its own, finds the events, and sends you a neatly organized table. It figures things out autonomously without anyone telling it each step. That's Agentic AI.

🧩 Agent = Someone Who Acts on Your Behalf

An "agent" means someone who acts on behalf of another. Just like a real estate agent finds a house for you, or a travel agent plans your trip — an AI agent does the work you entrust to it.

Here's a secret: pb has dozens of tools it uses when running errands. Checking the calendar, sending emails, searching the internet, creating files... pb decides on its own which tool to use!

💬 Chatbot vs. Agent: The Critical Difference

A chatbot has a simple input → output structure. If you ask "What's the weather in Paris?", it just returns text. It doesn't actually call a weather API, save the result, or decide what to do next.

An agent is different. It operates in a loop of think → act → observe → think again. When given a goal, it plans which tools to use, actually uses them, reviews the results, and decides the next step.

🔧 What Is a Tool?

The tools an AI agent can use are diverse: search engines, calendars, email, file systems, code executors, external APIs... Think of them like apps on a smartphone. The LLM brain decides which app to open and what to do with it.

Pebblous's NanoClaw (pb) has tools like Google Calendar, Gmail, Google Drive, and Ollama (a local AI model). So when you say "Draft this week's report," it checks the calendar for the schedule, finds related files on Drive, and creates the document itself.

📜 How AI Got Here

🔄 ReAct Pattern — Interleaving Thought and Action

ReAct, proposed in Google's 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models," has become the foundational paradigm for agent design. Rather than simply producing a final answer, the LLM explicitly generates a trace of Thought → Action → Observation.

Thought: The user wants a summary of this week's meetings. I need to check the calendar first.
Action: list_calendar_events(timeMin="2026-03-23", timeMax="2026-03-27")
Observation: [{"title": "Team Standup", "time": "Mon 10:00"}, {"title": "Investor Meeting", "time": "Wed 14:00"}]
Thought: There are 2 meetings. Now I need to get the details of each meeting.
Action: get_event_details(eventId="abc123")
Observation: {"description": "Q1 progress sharing", "attendees": [...]}
Thought: I have enough information. I'll write the summary.
Final Answer: 2 meetings this week — ...

The advantage of this approach is that the reasoning process becomes visible. You can debug what the agent was thinking and why it chose a particular tool.

🛠️ Function Calling and Tool Use

Modern LLM APIs define tool lists as JSON Schemas and pass them to the model. In certain situations, the model outputs structured function calls instead of text, and the execution environment converts these into actual API calls, feeding the results back to the model.

🔌 MCP — Model Context Protocol

MCP (Model Context Protocol), open-sourced by Anthropic in November 2024, is often compared to "the USB-C standard for connecting AI models to external tools." Previously, each agent framework had its own way of integrating tools, but MCP standardized this.

An MCP server exposes three core capabilities — resources (read), tools (execute), and prompts (reusable templates) — through a standard JSON-RPC interface. Since the client (agent) communicates with servers via the standard protocol, any MCP server can be swapped in.

Pebblous's NanoClaw operates by combining Google Calendar, Gmail, Google Drive, and Ollama MCP servers. OpenClaw (based on Claude Code) performs code tasks using filesystem, bash, and browser MCP servers.

🤝 Multi-Agent — Agents Working Together

For complex tasks, a multi-agent architecture where several agents divide and conquer is more effective than a single agent handling everything. An orchestrator agent breaks down the overall task and delegates to sub-agents.

The two agents can also collaborate as a team. When OpenClaw writes a blog post, NanoClaw sends notifications to the team via Slack/WhatsApp, and OpenClaw creates a PR — a seamless workflow.

📜 The History of Agent AI — Since the 1950s

🚀 Why Now — Five Convergences

The practical emergence of Agentic AI from 2023 onward resulted from the convergence of five technological and economic conditions.

🧠 Agency Delegation: The Critical Difference from Automation

The Principal-Agent Problem in economics addresses information asymmetry and incentive misalignment in delegation relationships. When viewing an AI agent as the "agent," this problem takes on an entirely new dimension.

Automation is vulnerable to exceptions. When a script receives unexpected input, it fails or behaves incorrectly. An agent is different — it maintains its goal even in unexpected situations and tries alternative approaches. This is the essence of Agency (autonomous capacity for action).

However, Agency delegation also entails new risks. An agent may cause unintended side effects while pursuing its goal (the instrumental goal alignment problem). Approaches like Anthropic's Constitutional AI attempt to constrain an agent's behavior within defined principles.

🦾 OpenClaw and NanoClaw — Pebblous's Agents

The two agents can collaborate as a single team. OpenClaw can serve as an orchestrator, creating and coordinating sub-agents, or NanoClaw can delegate specific code tasks to OpenClaw. This is what multi-agent architecture looks like when implemented in a real production environment.