Learn Agent Architecture

What Learn Agent Architecture does

The repository argues that the model reasons and the harness turns that reasoning into controlled action by running tools, keeping state across calls, gating side effects and coordinating loops. It explains the harness section by section: loop, tools, memory, permissions, context, tasks and interfaces. Every section uses the same four-part lens of opening, mechanism, per-system examples and failure modes.

Hermes Agent is described as a long-term assistant that remembers you, learns workflows and runs anywhere. It is read for memory, skills and always-on channels, and it is the worked example for sections 7, 9, 14, 16, 19, 21 and 22, at version v2026.7.1. Claude Code covers every section. Runnable sections include a src/loop.py and a demo.py. The README is available in English, Traditional Chinese, Simplified Chinese, Japanese and Korean, and three companion repositories cover memory, deepseek-harness and eval sets.

Key features

• Sections numbered 0 to 23 across eight layers, from the basic loop to a harness that runs itself
• Hermes Agent studied for memory, skills and always-on channels
• Four systems studied: Claude Code, Hermes Agent, mini-swe-agent and deepseek-harness
• Runnable sections with a src/loop.py and a demo.py
• README in English, Traditional Chinese, Simplified Chinese, Japanese and Korean
• Companion repos learn-agent-memory, learn-deepseek-harness and EvalGrill

When to use it

• Understanding how Hermes Agent handles memory, skills and messaging channels compared with other agents
• Learning the agent loop before building your own agent
• Comparing harness design choices across Claude Code, Hermes Agent and mini-swe-agent

Who it is for: Developers who want to understand how agent harnesses work before building or choosing one.

How it fits with Hermes Agent

Uses Hermes Agent as one of its worked examples for the sections on memory, skills and channels. It is a learning resource and not software that runs with Hermes Agent.

FAQ