Where Agent Agency Actually Comes From

What it does

This repository is a structured course in agent “harness engineering”—the loops, tool runtimes, sandboxes, memory, and coordination layers that wrap an LLM and give it agency. It splits the topic into 24 self-contained sections, each built around an opening, a mechanism, how real systems implement it, and failure modes. The accompanying Python code evolves section by section; diffing adjacent folders reveals exactly one new mechanism at a time.

The interesting bit

The project treats the harness as the star of the show, arguing that most agents share the same small model loop and differ only in the engineering around it. It studies three real systems—Claude Code, Hermes Agent, and mini-swe-agent—as worked examples, so you are reading production patterns rather than toy illustrations. The pedagogy is deliberately cumulative: by section 23 you have built up to graph engineering and evaluation harnesses from a bare messages[] loop.

Key highlights

• 24 sections across eight layers, from core loops and permission sandboxes to multi-agent coordination and graph engineering.
• Each section follows the same four-part lens: opening, mechanism, per-system implementation, and failure modes.
• Runnable src/ code with offline test.py checks and live demo.py runs against the Anthropic API.
• Real-system case studies: Claude Code (full harness), Hermes Agent (memory and skills), and mini-swe-agent (minimal complete loop).
• Companion repo learn-agent-memory continues the deep dive on memory scaling beyond section 9.

Caveats

• Live demos require an Anthropic API key; the setup does not mention other model providers.
• This is explicitly a learning resource and reference implementation, not a reusable library or framework.

Verdict

Bookmark this if you are building agent infrastructure and want to see how layers like tool dispatch, context management, and error recovery fit together in real code. Look elsewhere if you need a drop-in agent framework or a survey of model fine-tuning.