ChengYanSuo
C

Harness Engineering

The engineering methodology for getting an Agent to finish long tasks reliably. Built on walkinglabs' Harness Engineering, fused with practices forged in real projects.

This module answers the last question: how do you get an Agent to reliably finish a long task that takes many steps across many sessions?

The core thesis — model capability is a constant; whether it delivers is a function of the harness. The same model fails bare and succeeds wrapped in a full harness; success rates climb from ~20% to near-perfect without changing a single line of the model. This module uses walkinglabs' Harness Engineering as the skeleton, fused with a battle-tested methodology from real projects: features.json as the single source of truth, the STATUS + milestone triad, linear slicing, fixtures before code, the four-layer defense system, and context-isolated sub-agents.

Lecture order: set the target (why capable Agents still fail) → give the definition (what a harness actually is) → unfold each subsystem (repo as source of truth / instruction splitting / cross-session continuity / boundaries and the feature list / end-to-end verification and clean handoff).

For: engineers and PMs who need an Agent to do real work on long tasks — in fact, the course you're reading was itself built with this methodology.

Posts

C · 0 Harness · Post 0

Why Capable Agents Still Fail

Most failures aren't in the model weights, but in the layer of engineering infrastructure outside the weights — the thing we call the harness.

8 min read
C · 1 Harness · Post 1

Same Model, Wildly Different Results — The Harness Is the Real Variable

The Harness is all of the engineering infrastructure outside the model weights — the Five Subsystems are its components, the Four Layers of Defense are how you ship it.

9 min read
C · 2 Harness · Post 2

Why the Repo Must Be the Agent's Single Source of Truth

Information that isn't in the repo effectively doesn't exist for an Agent. Knowledge must be committed to files, not scattered across tools.

8 min read
C · 3 Harness · Post 3

Why One Giant Instruction File Drags Down Your Agent

A 600-line CLAUDE.md is a self-reinforcing failure loop; the cure is splitting it into an entry file plus topic modules.

7 min read
C · 4 Harness · Post 4

Why Long Tasks Lose the Thread Across Sessions

A long task is bound to exhaust its context; initialization plus persisted artifacts let a fresh session take over in three minutes.

8 min read
C · 5 Harness · Post 5

"Why Agents Overreach Yet Finish Nothing: WIP=1 and the Feature-List Primitive"

Running many tasks at once means all of them fail; lock the boundaries with WIP=1 plus feature primitives backed by executable verification.

8 min read
C · 6 Harness · Post 6

Why a Task Only Counts as Done After End-to-End Passes, and Why Every Session Must Leave a Clean State

Unit tests passing ≠ task done. The done-decision must be externalized by running E2E for real, and every session must close out against a five-dimension clean state to fight entropy.

9 min read
C · 7 Harness · Post 7

Why Multi-Agent Orchestration Should Be a Script, Not Model Improvisation

The quality of multi-agent collaboration comes from a hard-coded blueprint — where to parallelize, where to barrier, how many votes to verify, when to exit early. We dissect that blueprint using Claude Code's Workflow tool.

8 min read