Chapter 3: Three Paradigm Shifts#

You might ask: since models are so smart, why do we still need Harness?

The answer is: the model’s “smartness” has conditions. The better information you give it, the better it performs. The more stable control you give it, the more reliable its output.

Over the past three years, the industry has gone through three stages around the question “how to make models truly useful.”

3.1 Prompt Engineering (2023-2024): Learning to Talk to AI#

Starting with a Counterintuitive Phenomenon#

In early 2023, when ChatGPT first went viral, many people discovered a strange phenomenon: same model, ask it differently, answer quality varies wildly.

You ask AI “help me write a proposal,” it might give you something very generic. But if you say “you are a product manager with 10 years of experience, please help me write a competitor analysis proposal for a short-video app targeting young users, including market status, comparison of strengths and weaknesses of three main competitors, and differentiation suggestions” — the effect is completely different.

Why is that? Because the model generates answers based on the context you give it. The more ambiguous the context, the more it can only give you generic answers. The more specific the context, the more it can focus on what you actually want.

Techniques That Have Been Validated#

Thus, a new direction emerged: Prompt Engineering. Dedicated to researching how to write prompts that make AI output better results.

Several core techniques:

Role setting. Tell the AI who it is. “You are a senior Python developer” gets more professional code than “help me write code.” Why? Because role setting activates knowledge in the model related to this role.

Few-shot examples. Show the AI a few examples and let it imitate. You want the AI to extract article summaries; rather than explaining what a summary is for half a day, just give three examples of “original text → summary.” The model will automatically learn the pattern from the examples.

Chain-of-Thought. Let the AI think step by step. The phrase “please reason step by step” can improve accuracy on math and logic problems quite a bit. The reason is it forces the model to write out the intermediate reasoning process, rather than jumping directly to the answer.

Output format control. Specify whether you want JSON, tables, or Markdown. The clearer the format, the more stable the output.

A Key Realization#

Anthropic has a conclusion in their internal Prompt Engineering Workshop: the essence of Prompt Engineering is iterative debugging, not finding “magic words.”

There is no universal Prompt template. Good Prompts are tried out — write one version, look at the results, revise one version, look at the results again. This process is not fundamentally different from debugging code.

How to implement: In 2026, you don’t need to tune Prompts from scratch. Claude Code and Cursor both have built-in Prompt optimization features — you describe what you want, and it helps you generate and optimize Prompts. If you want to systematically manage Prompts, LangSmith provides Prompt version management and A/B testing features.

Where Are the Limitations#

But Prompt Engineering has a fundamental limitation: it only manages “how to speak,” not “what to see.”

No matter how good you are at asking questions, if the AI doesn’t know your company’s business background, doesn’t know the current project status, doesn’t know the latest market data, its answers can only be based on general knowledge learned during training.

Here’s an analogy: Prompt Engineering teaches you how to ask questions to a librarian — the more precisely you ask, the more correct the book you get. But whether the library has the specific book you want, that’s not something asking skills can solve.

This is why the next paradigm came.

3.2 Context Engineering (2025): What Information to Show AI#

The Essence of the Problem#

In 2025, the industry realized a key problem: the model’s bottleneck is not IQ, but information.

GPT-4, Claude, Gemini — these models’ reasoning capabilities are already very strong. When they don’t answer questions well, it’s often not because they’re dumb, but because they don’t know — don’t know your company’s background, don’t know what the user said before, don’t know the latest data.

Here’s a real scenario: you ask the AI to help you analyze competitors; the analysis it gives you might be very professional, Thus, a new concept was proposed: Context Engineering. The core idea is: rather than optimizing what you say to the AI, optimize what information the AI can see.

Five-Layer Information Stack#

Swyx (Latent Space host) proposed a practical framework: AI can see context in five layers.

First layer: System instructions. Tell the AI its role, rules, limitations. This is the most stable layer, usually set before the conversation starts.

Second layer: Conversation history. Messages between the user and AI. This layer lets the AI “remember” what was discussed before.

Third layer: Retrieved knowledge. Relevant information retrieved from external databases, documents, web pages. This layer is the core of RAG — the AI doesn’t need to remember all knowledge, only needs to look it up when needed.

Fourth layer: Working state. Progress of the current task, intermediate results, user preferences. This layer lets the AI handle multi-step tasks.

Fifth layer: Tool definitions. Which tools the AI can call, how each tool is used. This layer determines the AI’s action capability.

The five layers combined together are all the information the AI can see during each reasoning.

Core Technologies#

Context engineering has several key technologies, each solving a specific problem:

Context offloading. Store information outside, call it in when needed. Rather than stuffing all information into the context window, use RAG to retrieve on demand. This way, even with a limited context window, the AI can access vast amounts of knowledge.

Context compression. When conversations get too long, create summaries of historical messages. Keep key information, discard redundant details. This way, you can accommodate longer conversations within a limited window.

Prefix Caching. Cache unchanged system instructions and commonly used prefixes, avoiding recalculation each time. This can reduce latency and costs quite a bit.

An Important Discovery#

Anthropic discovered in 2025: many tasks that previously needed complex RAG pipelines can now be handled with direct file system access.

Give the Agent a grep tool and a read tool, let it go find the information it needs in the file system itself. This is more flexible and more reliable than pre-designing complex retrieval pipelines.

Behind this is the improvement in model capability — models have become stronger, don’t need such fine-grained preprocessing, they themselves can find the needed information from raw data.

Limitations#

Context engineering solved the “what to show AI” problem, but two problems remain unsolved:

First, the AI’s action capability. The AI can see information, Second, the AI’s safety boundaries. The AI can call tools, So, the third paradigm arrived.

3.3 Harness Engineering (2026): Building Control Systems#

Starting with a Failure Case#

A team built an AI customer service Agent. It performed very well in the Demo — whatever the user asked, it could give a professional answer. The team was excited and ready to launch.

On the first day of launch, problems came. A user asked a question the Agent wasn’t certain about; it didn’t say “I’m not sure,” but fabricated a seemingly professional wrong answer. The user operated based on this answer and caused losses.

The team’s post-mortem found: the problem wasn’t in the model, but in control. The model didn’t know when it should say “I’m not sure,” because no one told it this rule at the system level.

This is what Harness Engineering solves.

What is Harness#

Harness is a control system built around the model. It includes six subsystems: orchestration, tools, memory, sandbox, state, safety.

The model is the brain; Harness is the nervous system + skeleton + muscles. The brain is responsible for thinking,

Why 2026 is the First Year of Harness Engineering#

Models are strong enough now. Models in 2023 frequently made low-level mistakes; no matter how good the Harness, you couldn’t save them. By 2026, models have become reliable enough to handle complex multi-step tasks. Only then does the value of Harness truly manifest.

Tools are standardized. MCP (Model Context Protocol) became the industry standard for Agent-tool connections, donated to the Linux Foundation, with thousands of active Servers, and full support from Claude, Cursor, VS Code, Google Gemini. A2A (Agent-to-Agent) protocol became the standard for communication between Agents. Standardization lowered the threshold for building Harness.

Enough failure cases. 88% of enterprise Agent projects are stuck in production environments. The industry has accumulated enough failure lessons, knowing what should be done and what shouldn’t.

Why Harness is the Core Competitive Advantage#

One obvious change in 2026 is: the capabilities of mainstream models are converging.

GPT-5.5, Claude 4, Gemini 3, DeepSeek V4 — the performance gap between these models on most tasks is already small. Choosing which model depends more on price, speed, availability, rather than IQ.

DeepSeek V4-Flash’s price is $0.14/million input tokens, dozens of times cheaper than GPT-5.5, yet the performance gap on most tasks is less than 10%. In this situation, competitive advantage no longer comes from “which model you used,” but from “what you built around the model.”

Three years ago, the core work of building an AI product was model selection and Prompt tuning. Today, the core work has become designing orchestration logic, integrating tools, managing state, and ensuring security. These tasks have higher technical content and are harder to replicate.

Imagine an Agent as a car. The Model is the engine. In 2023, the engine didn’t have enough horsepower; no matter how you optimized the chassis, it was useless. By 2026, the engines are already strong enough — everyone’s horsepower is about the same. Harness is the transmission, chassis, suspension, brakes, steering system. When engines are about the same, the quality of the car depends on these supporting systems.

Where Are the 88% of Enterprise Agent Projects Stuck#

Demo vs Production. Demo environment: carefully prepared inputs, simple tasks, error-tolerant audience. Production environment: ever-changing user inputs, complex edge cases, zero-tolerance errors.

Six Major Failure Reasons:

1. No observability. Agents run in a black box; when problems occur, they can’t be diagnosed.
2. No retry and timeout mechanisms. Tool calls fail and then hang.
3. No cost control. 10,000 calls per day, costs run away.
4. No safety guardrails. The Agent deleted files it shouldn’t have, sent emails it shouldn’t have.
5. No evaluation system. Don’t know if the Agent is doing well or not.
6. Harness too primitive. Treating Agents as “model + Prompt,” ignoring other subsystems.

Chapter Summary#

Three years, three paradigm shifts:

Each generation is built on top of the previous one. Harness Engineering doesn’t deny Prompt and Context, but incorporates them as part of a larger system.

Starting from the next chapter, we’ll disassemble each of the six subsystems of Harness one by one. We’ll start with orchestration — it’s the Agent’s execution flow, and also the skeleton of the entire Harness.