← Blog

Harness Engineering: Implementation Models

July 16, 2026 AIAgentsHarness Engineering
Harness Engineering · Part 5 of 7
This is part of Harness Engineering , a living series I'm writing as I learn in public. It stays in draft on purpose: I update these posts as I go deeper, so expect it to grow and change.

You do not build all of this at once. The paper lays out a ladder, H0 to H3, where each rung adds exactly one class of runtime support. Building it this way is also how you measure it: each level is a controlled ablation that exposes one named kind of help, so you can see what it buys you.

H0 · Minimaltask + repo files H1 · Tools+ tool & test registry H2 · Context & Memory+ project memory, task state H3 · Observe & Verify+ attribution, verification
  • H0, minimal: the task description and the repository files, nothing else.
  • H1, tool harness: adds a tool registry, a test-command registry, and a protocol for using them.
  • H2, context and memory: adds project memory, a task-state file, and a context-selection protocol.
  • H3, observe and verify: adds a deterministic check registry, a bug-reproduction protocol, failure attribution, and a verification protocol.

At H3 the agent stops guessing and starts working like an engineer: reproduce, attribute, fix, verify, report. That loop is the difference between an agent that says it fixed the bug and one that can show you.