agent-harness-course

KO | EN

gitlite — search

#javascript #python #ai #typescript #llm #golang #go #machine-learning #deep-learning #linux #awesome #vue #mcp #ios

agent-harness-course

(★ 21)

A hands-on course for building an AI agent harness from scratch — grows one project from a single LLM call to tool calling, memory management, planning, guardrails, tracing, and multi-agent orchestration. No frameworks.

agent-harness-course 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Building an Agent Harness — A Hands-On Course

This course teaches you how an "AI agent" is really built, one concept at
a time, by growing a single project from a 1-turn LLM call into a
production-shaped agent harness with tools, memory management, planning,
error recovery, guardrails, observability, multi-agent orchestration, and
a packaged reusable library.

## Who this is for

You already know Python and have called an LLM API before (OpenAI,
Anthropic, etc.). You don't need any prior "agent framework" experience —
in fact, this course deliberately avoids frameworks like LangChain/AutoGen
so you build the underlying mechanics yourself and understand exactly what
those frameworks are doing under the hood.

## Philosophy

Every module **builds on concepts from the previous modules**. By the end,
you will have watched one project grow from ~40 lines to a multi-file core
package, with advanced policies and workflows demonstrated in focused
standalone modules. You'll understand each piece because you built it
incrementally rather than being handed a finished framework.

All code uses the **OpenAI-compatible API format**, via the official
`openai` Python SDK. The examples are tested with Ollama's compatible
endpoint and use the same request shape as:
- OpenAI's real API
- A local [Ollama](https://ollama.com) server (`ollama serve`, OpenAI-compat
  mode at `http://localhost:11434/v1`)
- Other OpenAI-compatible providers (vLLM, LM Studio, together.ai, etc.)

You switch between the default Ollama setup and OpenAI by changing three
constants (`BASE_URL`, `API_KEY`, `MODEL`) at the top of each file. Exact
feature and parameter support varies among other compatible providers, so
check the provider's compatibility documentation when using one.

## Setup

Create and activate a project-local virtual environment, then install the
dependency inside it:

macOS/Linux:

```bash
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install openai
```

Windows PowerShell:

```powershell
py -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install openai
```

Run the module commands below while the virtual environment is activated.
When finished, leave it with `deactivate`.

If using Ollama locally:
```bash
ollama pull gemma4:e4b   # recommended default for this course
ollama serve
```

`gemma4:e4b` is the recommended default for this course. With the default
Ollama configuration, the examples send `reasoning_effort="none"` so the
agent-loop and tool-calling behavior stays easy to observe while learning.
That provider-specific option is omitted after replacing the dummy `ollama`
API key with a real provider key, because not every model supports it.

Whichever variant you choose, set the same model name in each module's
`MODEL` constant.

If using real OpenAI, set `BASE_URL = "https://api.openai.com/v1"` and
`API_KEY` to your real key in the config section of whichever module
you're running.

Every module script is runnable standalone:
```bash
python module00/module00_baseline.py
python module01/module01_agent_loop.py
# ...etc
```

## Module Index

| Module | Topic | What gets added |
|---|---|---|
| [0](module00/README.md) | Why a Harness? | Baseline: a single LLM call, no loop, no tools |
| [1](module01/README.md) | Minimal Agent Loop | Multi-turn loop + conversation history |
| [2](module02/README.md) | Your First Tool | Real tool/function calling (a calculator) |
| [3](module03/README.md) | Tool Registry & Multiple Tools | Decorator-based registry, robust error handling |
| [4](module04/README.md) | State & Memory Management | History compression via LLM-generated summaries |
| [5](module05/README.md) | Planning & Task Decomposition | Plan → Execute → Synthesize pattern |
| [6](module06/README.md) | Error Recovery & Self-Correction | Retry limits + reflection/critique loop |
| [7](module07/README.md) | Parallel Execution & Performance | Async tool calls, rate limiting |
| [8](module08/README.md) | Guardrails & Safety | Human-in-the-loop confirmation, input validation, audit log |
| [9](module09/README.md) | Observability & Debugging | Structured tracing, saved JSON traces, trace viewer |
| [10](module10/README.md) | Multi-Agent & Orchestration | Sub-agents-as-tools pattern (researcher + writer) |
| [11](module11/README.md) | Practical Deployment | Packaged reusable library: cost tracking, caching, prompt versioning |
| [12](module12/README.md) | Production Roadmap | What to study and harden after completing the course |

## Files in this project

```
module00/                     # Modules 0-10 each contain a script and README
  README.md
  module00_baseline.py
module01/
  README.md
  module01_agent_loop.py
# ... module02/ through module10/ follow the same pattern
module11/
  README.md
  example_usage.py
  agent_harness/              # packaged reusable library
    __init__.py
    config.py
    tools.py
    prompts.py
    cost.py
    cache.py
    loop.py
module12/
  README.md                   # roadmap from educational code to production
README.md                     # this file
```

## How to work through this course

1. Read a module's README (linked above) before running its code —
   each one explains the new concept and exactly what changed from the
   previous module.
2. Run the script and watch the step-by-step console output — every
   module prints what it's doing at each step so you can follow the
   mechanics live, not just read about them.
3. Try the "Things to try" section at the end of each module README —
   small modifications that deepen your understanding.
4. Move to the next module once you're comfortable with the current one's
   code — remember, the next module's diff will be small; most of the
   file will look familiar.

By Module 11, you'll have a small reusable core library
(`module11/agent_harness/`) that a new project could `import` directly
instead of copy-pasting a script. Modules 4-10 remain focused examples of
advanced features you can compose around that core.

## Caveat: Educational Use Only

This project is intentionally designed for teaching. The examples favor
clarity and incremental learning over production-grade security,
reliability, performance, and optimization. Some implementations are
deliberately simplified, including tool validation, permission controls,
error handling, retries, caching, cost tracking, logging, and secret
management.

Before adapting any part of this project for production, perform a full
security and architecture review. At a minimum, consider strict input and
output validation, sandboxing and least-privilege access, authentication
and authorization, secure secret storage, rate limits and timeouts,
failure recovery, persistent state, privacy and data retention,
observability, testing, dependency management, model behavior evaluation,
scalability, latency, and cost controls. The sample code should be treated
as a learning foundation, not as production-ready software.

## Fun Fact

I also used AI while developing this project.

I defined the goals, curriculum, architecture, module boundaries, and implementation constraints. 
AI helped accelerate coding, review, documentation, and refinement.

That process reinforced one of the central ideas behind the course:

AI can generate and improve individual pieces of work, but the surrounding system still needs clear goals, structure, constraints, evaluation, and human judgment.

The same principle applies when building agents. A capable model is only one component. Reliable behavior comes from how the complete system is designed around it.

agent-harness-course

agent-harness-course 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

agent-harness-course

agent-harness-course 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README