Code Assistant Example

Article

The article on Level Up Coding gives a hands-on overview of this example.

The directory containing the example files can be found here.

This example builds a simplified "Claude Code" - a terminal coding assistant powered by a mesh of three autonomous agents. It demonstrates how to compose specialized agents into a system where a brain reasons, hands execute, and a coordinator orchestrates - just like a real AI coding assistant.

It highlights:

How an orchestrator agent coordinates both LLM and tool agents
How LLM-to-LLM delegation works via agent_call with infer
How tool delegation works via agent_call with tool_call
How separation of concerns (brain vs hands) enables safer, more reliable systems
How agents discover each other dynamically through the Registry

💬 High-Level User Request

Add docstrings to all functions in utils.py

User Request

From this single input, the system:

Explores the workspace to understand the project structure
Reads the target file to get the current code
Reasons about what docstrings are needed (using the Planner's LLM)
Generates the complete updated file with proper PEP 257 docstrings
Writes the changes to disk
Summarizes what was changed for the user

🧩 Agent Overview

1. Orchestrator Agent (Coordinator | LLM + Agent Calls)

Role: User-facing coordinator and workflow manager.

Receives the user's coding request
Decides which agent to call, when, and with what context
Delegates reasoning to the Planner and file operations to the Coder
Aggregates results and presents a summary to the user

Uses an LLM for decision-making, but never touches files or generates code directly. It coordinates.

2. Planner Agent (Brain | LLM-Only)

Role: Code analysis, planning, and code generation.

Analyzes coding tasks and creates step-by-step implementation plans
Reviews existing source code for issues and improvements
Generates precise, complete file contents ready to be written
Responds with structured output following PEP 8 / PEP 257 conventions

No filesystem access. The Planner is a pure reasoning engine, it thinks, but cannot act. This is the infer delegation in action.

3. Coder Agent (Hands | Tools-Only)

Role: Deterministic filesystem operations.

read_file(path) - Read file contents with line numbers
write_file(path, content) - Create or overwrite files
list_directory(path) - List directory contents
search_in_files(pattern, path, file_filter) - Grep-like search

No LLM. The Coder executes file operations reliably and deterministically. All operations are sandboxed to a configurable workspace directory for safety.

🧠 Design Note: Why Three Agents?

This architecture mirrors how production coding assistants (Claude Code, Cursor, etc.) work internally:

Concern	Agent	Why Separate?
Reasoning	Planner	Can use a powerful (expensive) model; no side effects
Execution	Coder	Deterministic; can run on a secure file server
Coordination	Orchestrator	Can use a lighter (cheaper) model; just routes work

The key insight: the brain that generates code should not be the same component that writes files. Separation makes the system safer, more testable, and independently scalable.

🔁 Sequence Diagram

Two Types of `agent_call`

The Orchestrator uses both delegation modes available in Protolink:

Delegation	Target	Action	What Happens
LLM-to-LLM	Planner	`infer`	Orchestrator's LLM sends a prompt → Planner's LLM reasons → response returned
LLM-to-Tool	Coder	`tool_call`	Orchestrator's LLM specifies tool + args → Coder executes → result returned

This is the core of Protolink's agent mesh: agents delegating to each other over the network using a standardized protocol. The Orchestrator doesn't know (or care) whether the Planner and Coder are on the same machine or across the globe.

🧠 Agent Classification Summary

Agent	Uses LLM	Has Tools	Purpose
Orchestrator Agent	✅	❌	Coordination, routing, decision-making
Planner Agent	✅	❌	Code analysis, planning, code generation
Coder Agent	❌	✅	File read/write/list/search

🎯 Why This Example Matters

Relatable use case: Every developer understands what a coding assistant does
Both delegation modes: Demonstrates infer (LLM-to-LLM) and tool_call (LLM-to-Tool) in one system
Separation of concerns: Brain, Hands, and Coordinator are cleanly separated
LLM-agnostic: Switch between OpenAI, Anthropic, Ollama with a single environment variable
Safety by design: Workspace sandboxing prevents agents from accessing arbitrary files
Autonomous multi-step: The entire workflow runs without human intervention after the initial request

🧠 The Inference Loop - How the Orchestrator Works

When the Orchestrator receives a user request, Protolink runs an inference loop:

The Orchestrator's LLM reads the system prompt + discovered agent cards (from the Registry)
The LLM decides which agent to call and outputs a structured agent_call
Protolink intercepts the agent_call, resolves the agent URL via the Registry, and sends an HTTP request
The target agent processes the request and returns a result
The result is injected back into the LLM's conversation as an observation
The LLM decides the next step (another agent_call, or a final response)
Steps 2–6 repeat until the LLM produces a final response

┌─────────────────────────────────────────────────────────┐
│                   INFERENCE LOOP                        │
│                                                         │
│  ┌──────────┐    ┌─────────┐    ┌──────────────────┐   │
│  │ LLM      │───→│ Protolink│───→│ Remote Agent      │   │
│  │ thinks   │    │ routes   │    │ executes          │   │
│  └──────────┘    └─────────┘    └──────────────────┘   │
│       ▲                                    │            │
│       │              observation           │            │
│       └────────────────────────────────────┘            │
│                                                         │
│  Loop continues until LLM produces "final" response     │
└─────────────────────────────────────────────────────────┘

This is how a single Task.create_infer(prompt=...) can trigger an autonomous, multi-step workflow spanning multiple agents.

📁 Files and Structure

examples/code_assistant/
├── run.py                  # ⭐ Entry point, starts all agents & demo
├── orchestrator_agent.py   # LLM coordinator (delegates to Planner & Coder)
├── planner_agent.py        # LLM-only reasoning (code analysis & generation)
├── coder_agent.py          # Tools-only (read, write, list, search files)
├── .env.example            # Environment template
├── README.md               # Detailed README with setup instructions
└── workspace/              # Demo Python project (auto-created by run.py)
    ├── main.py             # Simple calculator app
    ├── utils.py            # Utility functions (without docstrings)
    └── config.py           # App configuration

🚀 Running the Example

Install Protolink
```
pip install "protolink[http,llms]"
```

Configure your LLM provider

=== "OpenAI"

export OPENAI_API_KEY=sk-...

=== "Anthropic"

export ANTHROPIC_API_KEY=sk-ant-...

=== "Ollama (Free, Local)"

# Install from https://ollama.ai
ollama pull gemma4:e4b
ollama serve

Run the demo

=== "OpenAI"

cd examples/code_assistant
LLM_PROVIDER=openai python run.py

=== "Anthropic"

cd examples/code_assistant
LLM_PROVIDER=anthropic python run.py

=== "Ollama"

cd examples/code_assistant
LLM_PROVIDER=ollama python run.py

Try different queries

python run.py "Add type hints to all functions in utils.py"
python run.py "Create a test file for utils.py"
python run.py "Refactor utils.py to use a Calculator class"

📋 Expected Output

🤖 CODE ASSISTANT - Protolink Multi-Agent Coding System
======================================================================

📂 Setting up demo workspace...
   Created main.py
   Created utils.py
   Created config.py

📡 Starting Registry...
   Registry running at http://localhost:9000

🔧 Starting Coder Agent (tools-only, no LLM)...
   Coder running at http://localhost:8030
   Tools: ['read_file', 'write_file', 'list_directory', 'search_in_files']

🧠 Starting Planner Agent (LLM: openai)...
   Planner running at http://localhost:8020

🎯 Starting Orchestrator Agent (LLM: openai)...
   Orchestrator running at http://localhost:8010

🔍 Verifying agent discovery...
   Discovered 3 agents:
   • orchestrator (LLM): reasoning
   • planner (LLM): reasoning
   • coder (Tools): ['read_file', 'write_file', 'list_directory', 'search_in_files']

📝 Processing: "Add docstrings to all functions in utils.py"
======================================================================

⏳ Orchestrator is working...

   📁 [coder] list_directory: .
   📖 [coder] read_file: utils.py
   🧠 [planner] infer called: Add docstrings to these functions...
   ✍️  [coder] write_file: utils.py

======================================================================
✅ RESULT:
----------------------------------------------------------------------
Added comprehensive docstrings to all 6 functions in utils.py:
• add(a, b) - Addition operation
• subtract(a, b) - Subtraction operation
• multiply(a, b) - Multiplication operation
• divide(a, b) - Division with zero-check
• power(base, exponent) - Exponentiation
• factorial(n) - Recursive factorial

All docstrings follow PEP 257 conventions.
----------------------------------------------------------------------

🎯 Protolink Features Showcased

#	Feature	Where
1	`agent_call` with `infer`	Orchestrator → Planner (LLM-to-LLM reasoning)
2	`agent_call` with `tool_call`	Orchestrator → Coder (file operations)
3	Registry Discovery	Agents find each other dynamically at runtime
4	LLM-Agnostic	One-line switch: `create_llm("openai")` → `create_llm("anthropic")`
5	Transport-Agnostic	All agents use HTTP; switchable to SSE, WebSocket, or runtime transports
6	Tool-Only Agents	Coder has tools but no LLM - pure determinism
7	LLM-Only Agents	Planner has LLM but no tools - pure reasoning
8	Custom `handle_task`	Planner subclasses Agent for observability
9	Autonomous Orchestration	Multi-step workflow without human intervention
10	Workspace Sandboxing	File operations constrained to safe directory

🧩 Extending the Example

Add a Reviewer agent (LLM-only) that reviews code changes before they're written
Add a Test Runner agent (tool-only) that runs pytest after edits
Switch transports - use WebSocket for real-time streaming of edit progress
Add MCP tools - import tools from external MCP servers (e.g., GitHub, Jira)
Use different LLMs per agent - cheap model for Orchestrator, powerful model for Planner

📚 See Also

Getting Started – Core concepts and setup
Agents – Agent lifecycle and tools
Transports – Switching between HTTP, WebSocket, and runtime transports
Tools – Native and MCP tool integration
LLMs – LLM backends and usage
Ticket Booking Example – Another multi-agent example

💬 High-Level User Request​

🧩 Agent Overview​

1. Orchestrator Agent (Coordinator | LLM + Agent Calls)​

2. Planner Agent (Brain | LLM-Only)​

3. Coder Agent (Hands | Tools-Only)​

🧠 Design Note: Why Three Agents?​

🔁 Sequence Diagram​

Two Types of agent_call​

🧠 Agent Classification Summary​

🎯 Why This Example Matters​

🧠 The Inference Loop - How the Orchestrator Works​

📁 Files and Structure​

🚀 Running the Example​

📋 Expected Output​

🎯 Protolink Features Showcased​

🧩 Extending the Example​

📚 See Also​