CHAPTER 17 · Building the Full Mini-Harness · 10 / 10
Key takeaways
- The capstone is the eleven MVPs wired together: a loop (Ch 2) that builds prompts (Ch 3, 9, 10), manages context and budget (Ch 4, 5, 12), and runs tools (Ch 6) through an approval gate (Ch 7), in parallel (Ch 8), with optional subagents (Ch 11).
- There is exactly one seam to a real model (
call_model); swapping it turns the demo into a working agent, proving Chapter 1's point that the harness is the bulk of the work. - Safety is structural: every tool call passes a hook and an approval policy before executing, and a budget caps runaway loops.
- The architecture is small enough to read and faithful enough to scale; the production systems differ in polish and robustness, not in shape.
- Harden it next with real sandboxing, then extend with CodeAct, MCP, a two-tier review, observability, and research mode.