"Harness Engineering Basics (1/4) — Why the Work Environment Matters More Than the Model"

If a strong model still produces uneven results, the bottleneck is often not the model. It is the environment around it: what it reads, which tools it can call, how failures are detected, and how long work is handed off. That full environment is what this series calls the harness.

Key Takeaways

  • The practical gap between AI agent teams is increasingly less about model choice and more about the work environment wrapped around the model.
  • A harness is broader than a prompt. It includes instruction files, context assembly, tool surface, permissions, verification loops, logs, and handoff artifacts.
  • Across our research notes, better agents usually did not win because the model was smarter. They won because the system made mistakes harder to commit and easier to detect.
  • That is why harness engineering is better understood as a higher-level systems problem, not a subtopic of prompt writing.

1. Why people now say "the harness matters more than the model"

At first, model quality gaps were so large that upgrading the model often did move the needle. As agent usage matured, another pattern became harder to ignore. Teams using similar model families were getting very different results on similar work.

That gap usually appears outside the model itself.

  • what the model reads first
  • how tools are named and described
  • where failures are caught by tests or reviews
  • how long work is handed off across sessions

In other words, the model's workbench has become part of the product.

2. What a harness actually is

The word comes from physical systems that bind and connect parts together. A wiring harness in a car connects the engine, sensors, and power lines. An AI harness connects a model to the real world of files, tools, rules, and feedback.

In software terms, a raw model may be a powerful reasoning engine, but it does not reliably decide on its own what to read, what to execute, when to stop, or how success is judged. The harness fills that gap.

For this series, the working definition is simple:

Harness = the full work environment around the model

At minimum, that includes:

  • instruction structure that defines goals and boundaries
  • context design that shows only the necessary material
  • tool surface for reading, searching, and acting
  • permission boundaries such as approvals and sandboxing
  • verification loops such as tests, reviews, and retries
  • state-preservation artifacts such as handoffs, memory, and logs

3. How this differs from prompt engineering

Prompt engineering focuses on improving the wording of the model input. Harness engineering focuses on the full runtime in which the model repeatedly works.

The difference is easiest to see by scope.

Question Prompt engineering Harness engineering
Main unit one input/output exchange a multi-step work loop
Main concern wording, role, format context, tools, permissions, checks, records
Typical failure ambiguous instruction unstable system structure
Typical fix rewrite the prompt redesign the environment and feedback loop

That is why agent work often improves more from "where does the system catch mistakes?" than from "how do we make the prompt sound smarter?"

4. Three recurring patterns from our research

Using sources/260518_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_15์žฅ_๋ธ”๋กœ๊ทธํ™œ์šฉ๋…ธํŠธ.md and the series design document as the base, three patterns kept repeating.

4.1 Structure beats bulk instruction

Dumping long guidance into one message does not reliably help. A cleaner split across AGENTS.md, CLAUDE.md, skill docs, and handoff files works better because the model can find the right rule at the right time.

4.2 Tool surface matters more than tool count

Adding more tools is less useful than exposing tools with names, descriptions, and inputs the model can judge correctly. A bad tool surface turns powerful capabilities into repeated misuse.

4.3 Verification and handoff come before "memory"

Long-running work often gets framed as a memory problem, but stable operation usually needs progress files, handoff notes, tests, and logs first. Durable memory helps later. Clear recovery structure helps immediately.

5. In our repository, harness layers are already familiar

Harness engineering can sound abstract until you map it to concrete repository structures. In this workspace, the pieces are already visible.

Harness layer Repo-native example What it does
Instruction structure AGENTS.md, CLAUDE.md fixes roles, boundaries, output rules
Context map tasks/plan.md, docs/memory-map.md narrows what should be read now
Handoff structure tasks/handoffs/, tasks/sessions/ carries long work across sessions
Verification layer quality gate, review flow exposes errors early
Permission boundary publish restrictions, protected config/ blocks unsafe actions structurally

So a harness is not a futuristic concept. It is often the name for the operating rules you already depend on.

6. Why strong models still produce unstable outcomes

A better model does not automatically solve the operational problems below.

Context overload

If every document is injected at once, important rules get buried.

Tool misuse

Ambiguous tool names, wide permissions, and noisy outputs weaken model judgment.

Missing verification

Without tests or review loops, agents repeat the same mistakes.

Session discontinuity

Without progress and handoff artifacts, each new session has to reconstruct context from scratch.

All four are usually more tractable through harness design than through model shopping.

7. Where practitioners should start

If "harness engineering" stays too abstract, it becomes empty branding. For beginners, the practical sequence is simpler.

  1. Separate must-follow rules into a short instruction file.
  2. Distinguish always-needed context from occasionally-needed context.
  3. Clean up tool names, descriptions, and inputs to reduce misuse.
  4. Add at least one verification loop: tests, reviews, or checklists.
  5. For long tasks, write handoff artifacts before building fancy memory layers.

This is not glamorous, but it is where repeatability usually begins.

8. What the rest of this series will cover

This A1 article sets the definition and problem statement. The next entries move down one layer at a time.

  • A2: how AI agents actually work through context, tool calls, and the agent loop
  • A3: why instruction structure and context design matter more than longer prompts
  • A4: MCP, tool engineering, and why the tool surface must be designed deliberately

The goal is not to stop at "harness matters." It is to break down which layers matter, and how to design them.

References

  • docs/blog_series_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_์ด๊ด„_design.md
  • sources/260518_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_15์žฅ_๋ธ”๋กœ๊ทธํ™œ์šฉ๋…ธํŠธ.md
  • Martin Fowler, Harness engineering for coding agent users
  • WikiDocs, Chapter 1 notes from ํ•˜๋„ค์Šค ์—”์ง€๋‹ˆ์–ด๋ง ๋ฐฑ๊ณผ์‚ฌ์ „

This is Part 1 of the Harness Engineering Basics series. Next: How AI agents actually work — context, tool calls, and the agent loop.

Series overview: Harness Engineering Series Guide

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »