Agentic AI Harness
The diagram below maps out the key components of an agentic AI system — how an LLM, an orchestrating agent, tools, and a client all fit together into a working harness.
What the Diagram Shows
The system has four layers that interact to produce intelligent, multi-step behavior.
LLM — Reasoning & Planning At the top sits the Large Language Model. It is not the agent itself — it is the reasoning engine the agent calls on. The LLM interprets goals, decides which tools to invoke, and generates the responses that flow back down to the client.
Agent — Orchestrator The central component is the Agent. It manages three core responsibilities:
- Goal Decomposition — breaks a user request into a sequence of concrete sub-tasks
- Tool Selection — decides which system tool or capability is needed for each step
- Memory & State — tracks context, intermediate results, and progress across the full task lifecycle
System Tools (Environment) On the left are the tools the agent uses to interact with the local environment:
| Tool | Purpose |
|---|---|
| File System | Read, write, and list files and directories |
| Bash / Shell | Execute arbitrary shell commands |
| Git | Repository operations and version control |
| Other Tools | Process management, env vars, network, schedulers |
Tools (Capabilities) On the right are higher-level capability integrations:
| Capability | Purpose |
|---|---|
| Skills | Pre-built or custom functions the agent can invoke |
| MCP | Model Context Protocol servers for structured tool access |
| Plugins | Third-party service integrations |
| RAG | Retrieval Augmented Generation for grounding responses in external data |
Client — Interaction / I/O At the bottom, the client represents every surface through which a user or system can reach the agent: Chat UI, mobile app, desktop app, or API/SDK. The client sends requests and receives the agent's final responses and actions.
How It All Flows
A request arrives from the client. The agent receives it, calls the LLM to reason about the goal, decomposes the goal into steps, and dispatches tools from either the environment (left) or the capability layer (right) to complete each step. Results accumulate in the agent's memory and state, and the final response is returned to the client. This loop repeats until the goal is satisfied — which is what makes the system agentic rather than a simple one-shot prompt.