Tool Calling, MCP, and Skills: How Agents Use External Capabilities
A practical, implementation-level explanation of tool calling, MCP, skills, and why CLI-based agents have become common.
Shion Honda
I'm Shion Honda, a software engineer based in Rome since February 2026. At work I focus on turning AI agents into product features that people can actually rely on. That naturally pulled me toward evaluation, context engineering, and tool design: the messy layer between a model demo and a system that survives production.
On this site I write implementation-level essays about LLM behavior, agent infrastructure, and the trade-offs behind modern AI products, with occasional notes on software, books, and life in Europe. I also keep track of what I read on my Bibly bookshelf, a social media for readers I'm currently building. Before moving deeper into product engineering, I worked on machine learning systems at scale and studied molecular representation learning at the University of Tokyo; some of that earlier academic work is still collected on Google Scholar.
On this site I write implementation-level essays about LLM behavior, agent infrastructure, and the trade-offs behind modern AI products, with occasional notes on software, books, and life in Europe. I also keep track of what I read on my Bibly bookshelf, a social media for readers I'm currently building. Before moving deeper into product engineering, I worked on machine learning systems at scale and studied molecular representation learning at the University of Tokyo; some of that earlier academic work is still collected on Google Scholar.
What I work on
- Implementing AI agents as real product features, not isolated demos
- Evaluation, context engineering, and tool orchestration for LLM systems
- Writing technical essays that explain why systems behave the way they do
Selected writing
Why You Can’t Set Temperature on GPT-5/o3
Some LLMs disable sampling knobs like temperature and top_p. Here’s why.
Top 5 AI Papers of 2025
This article picks five notable AI papers from 2025 and summarizes their key ideas and limitations, including reinforcement learning for reasoning, agent benchmarks, long‑task metrics, and a statistical explanation of hallucinations.
Reuse and licensing
Unless otherwise noted, the written content and original images on this blog are licensed under CC BY 4.0. You may reuse them, including for commercial purposes, as long as you provide attribution with a link to the original page and indicate whether you made changes.
Source code in the repository is licensed under the MIT License.
If you cite a post from this blog in a paper, you can use the following BibTeX template and replace the placeholders.
Source code in the repository is licensed under the MIT License.
If you cite a post from this blog in a paper, you can use the following BibTeX template and replace the placeholders.
@online{honda_YEAR_slug,
author = {Shion Honda},
title = {POST_TITLE},
year = {YEAR},
month = {MONTH},
day = {DAY},
url = {https://hippocampus-garden.com/POST_SLUG/},
note = {Accessed: YYYY-MM-DD}
}Career
2026-
Software Engineer, Alan
Moved to Rome in February 2026 and continued building AI-agent product features, with a recent focus on evaluation and context engineering.
2023-2026
Software Engineer, Alan
Moved from Tokyo to Paris to join Alan and expanded from machine learning into broader product and backend engineering.
2020-2023
Machine Learning Engineer, Recruit
Worked on recommendation and machine-learning systems at production scale for consumer products.
2018-2020
University of Tokyo
Completed a master's degree in information science. Research focused on representation learning for molecules.
Books and external writing
Original title: 『[実践]データ活用システム開発ガイド』
Co-authored. Published in 2022.
Original title: 『ダイテル Pythonプログラミング』
Japanese translation. Published in 2021.
Technical writing for Alan and other external publications.
Older writing on topics such as graph neural networks, deep reinforcement learning, and generative models.