Building a General-Purpose LLM Agent from Scratch

A step-by-step tutorial that incrementally builds an autonomous LLM agent — from a simple chatbot to a system resembling Cursor or Claude Code.

Each step is a self-contained, independently runnable project. Later steps build upon earlier ones, adding one new core capability at a time.

Roadmap

Step	Name	Core Capability
01	Basic Chatbot	LLM API calls, multi-turn conversation
02	Context Management	Token counting, context inspection, debug view
03	Tool Use	Function calling, tool schema, smart compaction, calibrated token tracking
04	File Tools	File read/write/search + dynamic skill-based schema exposure
05	Shell Tool	Shell command execution with safety approvals
06	Error Recovery	Task-level retry, autonomous repair, optional failed-trace cleanup
07	CUDA Agent	Domain-specific CUDA kernel development with compile-verify-profile loop
08	Preemptible CUDA Agent	Queue-based interruptible autonomy with runtime task preemption
09	MCP Support	Model Context Protocol, dynamic tool discovery
10 (optional)	Streaming	Streaming output and richer terminal UX

Quick Start

cd 01_basic_chatbot
pip install -r requirements.txt

# Option A: Ksyun (mco-4 / opus4.6, used by default when key is set)
export KSYUN_API_KEY="your-key"

# Option B: InfiniAI (deepseek-v3, fallback when KSYUN_API_KEY is absent)
export INFINI_API_KEY="your-key"

python chatbot.py

Provider

The chatbot auto-detects the provider from environment variables — Ksyun takes priority when KSYUN_API_KEY is set, otherwise it falls back to InfiniAI.

Provider	Env var	Default model	Base URL
Ksyun	KSYUN_API_KEY	mco-4 (opus4.6)	https://kspmas.ksyun.com/v1/
InfiniAI	INFINI_API_KEY	deepseek-v3	https://cloud.infini-ai.com/maas/v1

Override the model at any time with --model <name>.

References

• Building Effective Agents — Anthropic
• InfiniAI API Docs