AI-Agents-Projects-Tutorials: A Notebook Collection, Not an Agent Framework

What this repository actually is

This is a tutorial index, not a library. Each README entry is a bullet with a title, a link to a notebook or Python file under Agentic AI Codes, AI Agents Codes, Agentic Workflows or Data Analysis, and a link to the matching MarkTechPost article. The topics list (agent-skills, agentic-rag, agentic-workflow, agentic-framework and similar) describes the subject matter of the notebooks, not the contents of an installed package. There is no setup.py, no pyproject.toml, no CLI entry point visible in the material. The primary language is Jupyter Notebook, which tells you the intended consumption mode: open a file, read the cells top to bottom, run them.

The audience is engineers and technical writers who want a working example of a named pattern. The Kimi CLI notebook demonstrates non-interactive agentic coding with JSONL streaming and session memory. The Omnigent notebook builds a policy-governed multi-agent financial research workflow. The OpenHarness notebook designs an agent runtime with tools, memory, permissions and multi-agent coordination. Each entry pairs a runnable artifact with an explanatory article, and the article is where the reasoning lives. The notebook is the proof that the pattern executes.

The pattern each notebook follows

The data flow is consistent across the entries that are described in the README. A notebook installs its dependencies in an early cell, defines or loads a model endpoint, registers tools or MCP connectors, runs an agent loop, and writes an artifact (traces, patches, chunks, SARIF reports, leaderboard tables). The DeepAnalyze-8B notebook is described as T4-friendly with sandboxed code execution and iterative analysis, which means the loop executes generated code inside a sandbox and feeds results back. The NVIDIA Open-SWE-Traces notebook parses trajectories and patches, then computes token budgets and tool-use metrics, producing supervised fine-tuning data. The Crawlee notebook builds a crawling pipeline with robots handling and exports RAG chunks.

What is missing from the material is the shared abstraction layer. There is no common base class or agent harness that these notebooks import. Each one appears to stand alone, which is fine for teaching and awkward for reuse. If you want the memory implementation from the nanobot notebook inside the OpenHarness runtime, you copy cells. The repository does not offer a composition story, and the README does not claim one.

Running a notebook: what the README gives you

The README does not include installation commands. What it gives you is a direct link per entry, for example the Kimi CLI notebook at Agentic AI Codes/Kimi_CLI_Advanced_Agentic_Coding_Marktechpost.ipynb and the Claude financial services script at Agentic AI Codes/claude_financial_services_agentic_workflows_marktechpost.py. The practical path is: clone the repository, open the target file, and read the first cells for the pip install lines and the environment variables the notebook expects. Several entries are aimed at Google Colab, which the titles state explicitly (the Fara browser agent tutorial and the nanobot tutorial both name Colab).

Because the material does not include the cell contents, I cannot list the exact package names or config keys. Treat the notebook's own setup cell as the source of truth. Two structural details do matter. First, the file extension varies: some entries are .ipynb and some are .py, so the execution path differs and a .py file will not have the cell-by-cell narrative. Second, model access is not uniform. The Fara tutorial is described as using a mock OpenAI-compatible endpoint, while other notebooks reference Claude, QwenPaw model providers, or DeepAnalyze-8B. Budget for API keys and for provider accounts before you start.

The licence field is empty and that is the first problem

The repository metadata shows License: (unknown). The README does not state a licence either. For a collection of notebooks whose purpose is to be copied into your own work, that is a real obstacle. You cannot tell from the supplied material whether you may reuse the code in a commercial product, whether attribution is required, or whether the notebooks carry a different licence from the surrounding text. The repository is not archived and the last push is dated 2026-09-04, so it is active, but activity does not resolve licensing.

If you plan to lift a cell into production, confirm the licence before you do. This is not legal advice, and the absence of a licence file is not the same as permission. It is a gap you have to close yourself, and the material offers no way to close it.

Where a notebook collection breaks down as a dependency

Tutorial notebooks rot. A cell that pins a model name, an SDK version or a hosted endpoint will fail once that endpoint changes, and nothing in this repository pins versions for you. The README is an index of articles, so the maintenance surface is the article links and the file paths. If a notebook moves, the bullet points at the old path. There are no releases retrieved, which means there is no changelog to check when something stops working.

The failure mode is predictable: you clone, open a notebook that looked relevant, and the first install cell resolves to a version that no longer matches the API calls below it. You then spend the session debugging the tutorial instead of learning the pattern. This is the wrong tool when you need a supported runtime with a version policy, an issue tracker with maintainer responses, or a dependency lock. It is the right tool when you want to see one specific pattern execute once, on a machine you control, with the article open beside it.

How it differs from an agent framework you install

Compare this to a framework such as LangGraph or CrewAI, which you add to requirements.txt and import. Those projects ship an abstraction (graphs, crews, state objects) that you configure, and their documentation assumes you are building an application rather than studying a technique. This repository does the opposite. It shows the technique at full resolution, including the parts a framework would hide: the raw JSONL stream from Kimi CLI, the trajectory parsing for Open-SWE traces, the SARIF output from SkillSpector, the lineage and low-cost reuse mechanism in OpenSpace.

The trade-off is explicit. A framework gives you a stable interface and takes on upgrade work. A notebook collection gives you visibility and takes on nothing. If you want to understand why an agent loop needs session memory before you adopt a library that provides it, the notebooks are the faster route. If you want to ship next week, they are not.

Maintenance cost and what to verify before you commit

The cost of using this repository is your own time, not a subscription. Each notebook carries its own dependency set, so the effort scales with how many patterns you pull in. If you reuse three notebooks, you inherit three sets of install cells and three model integrations, and you own the reconciliation. The repository's own maintenance appears to be adding new tutorial entries, since the README is a growing list dated across 2026. That is a content cadence, not a stability guarantee.

Verify three things before you build on a specific file. Open it and check whether it is a .ipynb or .py, because the README mixes both and the run instructions differ. Check which model or endpoint the notebook calls, since the Fara entry uses a mock OpenAI-compatible endpoint while the Claude entry assumes Claude access. Check the licence question above. If all three check out for your use case, the notebook is a reasonable starting point for that one pattern. If any of them is unclear, the article linked next to it is usually the better read, and it does not require you to run anything.

Editorial conclusion

Adopt this repository if you need a runnable reference for a specific pattern (Kimi CLI JSONL streaming, MCP connectors, OpenSpace skill lineage, SkillSpector scanning) and you are willing to read the linked MarkTechPost article before opening the notebook. Do not adopt it if you need a pinned dependency set, a supported runtime, or a licence you can put in front of legal. Before relying on any file, open it and check three things: the licence field, which model and endpoint each notebook calls, and whether the code is a .py script or a .ipynb notebook, because the README mixes both and the execution path differs.

Official sources

1. Issues
2. MARKTECHPOST-AI-MEDIA-INC/AI-Agents-Projects-Tutorials on GitHub
3. Project website
4. README