The vocabulary matters. The AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → is the decision loop of observe, inspect, choose, act; the agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → is the software scaffold around it that assembles prompts, exposes tools, tracks file state, and manages permissions.
Before doing anything the agent builds a small workspace summary of repo root, branch, git status, and files like AGENTS.md or README, because 'fix the tests' is not self-contained until you know which test command this project uses.
Rebuilding the whole prompt every turn throws away the provider's prompt cache. Split it into a stable prefix of rules, tool descriptions, and workspace facts, then append only the changing transcript and request.
Raschka's bet: drop a strong open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → model like GLM-5 into a Codex-quality harness and it would land near GPT-5.4, because vanilla model capabilities have converged and the scaffold now carries most of the difference.
Delegation is its own building block. Bounded subagents take a scoped task and return just the result, which keeps the main loop's context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → from filling with the details of a side errand.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Why it matters
If you want to understand how tools like Claude Code and Codex CLI actually work rather than just use them, this breaks the agent harness into six concrete building blocks and backs each with a minimal, readable Python implementation you can study end-to-end.
Key quotes
“the LLM is the engine, a reasoning model is a beefed-up engine (more powerful, but more expensive to use), and an agent harness helps us the model.”
“So the tricky design problem is not just how to spawn a subagent but also how to bind one :).”
