Capstone: Build a Production Harness

Module 12 · Master Course · 90 min

Build the whole thing from scratch. Every module applied. Your portfolio artifact.

The capstone

You've spent 11 modules learning the components. Now you integrate them into a working production harness — designed, built, audited, attacked, and defended by you.

This is the Course 1 credential: a portfolio artifact that proves you can do the whole job.

12.1 Design Document

12 decisions BEFORE code

#DecisionFrom
1Loop architectureModule 1
2Tool design (count, schemas)Module 2
3Context managementModule 3
4Memory tiersModule 4
5Sandboxing (inside/outside)Module 5
6Permission modelModule 6
7Error taxonomyModule 7
8CheckpointingModule 8
9Prompt assemblyModule 3/9
10SubagentsModule 1.3
11VerificationModule 9
12ObservabilityModule 10
Plus: the security threat model (Module 11) — which of the 9 defense layers in v1?

12.2 Build

Minimum viable production harness

  • ReAct loop + 5 stop conditions (Module 1)
  • 4 tools with schemas + structured errors (Module 2)
  • Compaction threshold (Module 3)
  • 2-tier memory: working files + episodic log (Module 4)
  • Sandbox: Docker, scoped fs/network (Module 5)
  • Risk-tiered permissions + capabilities (Module 6)
  • Error taxonomy: 4 categories (Module 7)
  • Atomic checkpointing (Module 8)
  • Computed verification (Module 9)
  • 8-field observability (Module 10)
  • Security: tagging + caps + scoping (Module 11)

12.3 Audit + Iterate

Self-audit + attack

6-phase methodology on your own harness (Module 0.3).
12-module scoring sheet — score yourself /60.
3 attacks from Module 11: indirect injection, memory poisoning, tool abuse. Fix what you find.
Architect's Verdict (3 sentences) + MLSecOps note (1 sentence) — for YOUR work.

Course 1 complete

  • You can design a harness — 12 decisions, defended.
  • You can build one — every module integrated.
  • You can audit one — 6 phases, /60 score.
  • You can attack one — 3 offensive techniques.
  • You can defend one — the 9-layer stack.
This is the Harness Engineer credential. Next: Course 2 — push this harness into offensive/defensive security. Or: the 21 deep-dives — apply the 6-phase method to real production harnesses.