Unofficial OpenClaw Architecture Study Notes

Slide 01

OpenClaw Orchestrator: AI Agent Functional Flow

Starting from the user's prompt, OpenClaw assembles context, injects tool rules, executes, routes, and persists memory before and after the LLM.

Slide 02

OpenClaw: The Stateful Orchestrator

OpenClaw Core sits between Tools, Memory, Users, and LLMs, orchestrating state, tools, and user interaction into a runnable agentic system.

Slide 03

The Proxy Misconception

OpenClaw does not simply forward prompts to the LLM. A proxy alone lacks durable state, execution capability, and the reason a real agent system exists.

Slide 04

The Black Box Isolation

The LLM is a black-box reasoning engine without long-term memory or physical agency. OpenClaw adds state and an execution environment.

Slide 05

The Standardized Contract

The LLM does not need to understand OpenClaw's internals. Both sides communicate through a standardized tool schema contract.

Slide 06

The 4-Step Asynchronous Protocol

The protocol covers Schema Injection, Tool Call Generation, Execution & Result, and Reasoning Continuation.

Slide 07

Pre-Flight Processing: High-Density Assembly

Before sending to the LLM, OpenClaw assembles persona, user info, memory, heartbeat, skills, tools, rules, and routing decisions.

Slide 08

Post-Flight Execution: Text to Physical Action

After the LLM outputs text, OpenClaw turns it into command execution, cross-platform reply routing, and memory updates.

Slide 09

The Core Orchestration Modules

Core modules include Observation, Planner, Policy Engine, and Event Bus, managing the full lifecycle of agentic state.

Slide 10

Diagnostic Matrix: Forwarder vs. Orchestrator

Compared to a simple forwarder, OpenClaw provides stateful management, tool orchestration, persistent memory, and proactive triggers.

Slide 11

The Super-Orchestrator

Through a standard protocol, OpenClaw gives the LLM long-term memory, environmental awareness, and executable "hands," making it more like an operating system for agentic AI.

Key Takeaways

OpenClaw is not a proxy but an orchestrator for state, tool orchestration, and memory.
The LLM only needs to return intent via a standardized schema contract; OpenClaw handles execution and persistence.
The system splits into pre-flight context assembly and post-flight execution/persistence, forming an agentic loop.
OpenClaw acts like an operating system layer for agentic AI—giving the model memory, environment, and executable "hands."