01 · Agent patterns¶
You are here: Agent patterns · Builder
Next: Harness engineering
This section is how an agent chooses its next step, and what it keeps. Harness engineering names the system around the model: what it can see, what it may do, and what gets checked. The other pages each press on one part of that system.
| Path | Builder |
| Time | About 95 minutes |
| You need | The agent loop |
Pages¶
| Page | You leave able to |
|---|---|
| Harness engineering | Separate the model from the six-part runtime, and match that runtime to the task |
| ReAct and plan-and-execute | Choose a tight loop or a written plan |
| Wiki memory | Design raw, wiki, and schema so the next run is smarter |
| How agents remember | Place working, episodic, semantic, and procedural memory, and write back |
If you only read one page in this section, read the harness. Wiki memory is how this repository itself remembers: search the guide, then write the answer back onto a page.
Practice¶
Stand up the folders in the research-agent playbook. Reading the pattern without a page on disk is still a chat.
Continue¶
Map | Paths | Glossary | References