从 Prompt 到 Context 再到 Harness：AI 工程的三次范式转移

In early 2026, Anthropic and OpenAI both published practical articles on Harness Engineering within almost the same week. Along with two academic papers on Agent memory infrastructure and community discussions on the evolution of three engineering paradigms, a complete picture is emerging.

The years 2023 to 2024 were the era of Prompt Engineering. The core challenge was how to talk to a model to get better responses. Wording, formatting, few-shot examples, Chain of Thought—all techniques revolved around a single conversation.

In 2025, Context Engineering became a mainstream concept. Shopify CEO Tobi's statement 'Context engineering is the new skill' was widely shared. The core issue shifted: prompting alone wasn't enough; the entire context window had to be engineered. RAG retrieval, long-context management, tool use orchestration, and memory systems all fell under this umbrella. You were optimizing all the information the model sees.

In early 2026, the concept of Harness Engineering was proposed almost simultaneously by two major companies. The core challenge escalated again: Agents could now run autonomously for hours or even days, so optimizing a single context window was far from sufficient. You needed to design the entire runtime environment for the Agent—including multi-agent collaboration architectures, evaluation feedback loops, mechanized enforcement of architectural constraints, and memory governance and verification mechanisms. The relationship among the three paradigms: each encompasses the previous one. Harness includes Context, Context includes Prompt. But the core problems they solve are fundamentally different.

Anthropic engineer Prithvi Rajasekaran's experiment revealed a counterintuitive fact: an Agent evaluating its own work is basically useless. Regardless of output quality, an Agent always gives itself positive ratings. The effect changed completely once generation and evaluation were split into two independent Agents. The evaluator didn't just read code and score—it used Playwright to actually interact with the page, clicking buttons, filling forms, and verifying functionality, then scored on four dimensions: design quality, originality, craftsmanship, and functional completeness.

In the frontend design experiment, after 5 to 15 rounds of back-and-forth between the generator and the evaluator, the generator produced a 3D spatial navigation solution around the tenth round.

The full-stack development experiment was more complex, with three Agents taking on separate roles: a planner expanded a one-sentence requirement into a full product specification, a generator implemented it step by step using React+FastAPI+PostgreSQL, and an evaluator performed QA testing.

The comparison data was clear. A single Agent in 20 minutes cost $9 and produced unusable output. A full harness in 6 hours cost $200 and delivered a complete game with sprite animations, AI integration, and export functionality. The most valuable finding: as Opus 4.6 improved, sprint decomposition could be removed, but the evaluator could not. Every component of the Harness encodes assumptions about model limitations. As models get stronger, some assumptions no longer hold, but some always will. Identifying which to keep and which to discard is a core skill of harness engineering.

OpenAI's experiment was more aggressive. In five months, a small team used Codex Agent to build a production system of roughly one million lines of code. Zero hand-written. Application logic, documentation, CI configuration, observability infrastructure, toolchains—all generated by the Agent. The role of engineers completely changed. They did three things: design the development environment, express intent through structured prompts, and provide feedback loops to the Agent. OpenAI called this 'depth-first working': breaking down large goals into small building blocks, letting the Agent construct each block, and then using those blocks to unlock more complex tasks.

Architecture governance was key to making this system work. Dependencies were organized into six strict layers: Types, Config, Repo, Service, Runtime, UI. Boundaries between layers were enforced mechanically by linters and CI—not by documentation conventions, but by code enforcement. PRs from the Agent that violated architectural constraints were automatically rejected.

Martin Fowler's comment is spot-on: Harness Engineering encodes context engineering, architectural constraints, and garbage collection into machine-readable artifacts that Agents can systematically execute.

Anthropic discussed evaluation loops, OpenAI discussed architectural constraints, but neither delved deeply into memory. That gap was exactly filled by two academic papers. The first, the (S)AGE paper, proposed a Byzantine fault-tolerant multi-agent memory infrastructure. The core question: when multiple Agents share a knowledge base, how do you ensure the knowledge written is trustworthy? An Agent could write false information due to hallucination, or could be attacked adversarially with fake memory injection.

Their solution is the Proof of Experience consensus mechanism. Each Agent has a reputation weight determined by four factors: historical accuracy, domain relevance, activity level, and number of independent verifications. Any memory an Agent submits must pass a weighted vote before being written to the knowledge base. Deployed on a 4-node BFT network, it achieved 956 req/s writes and 21.6ms P95 queries. Agents with this memory system had calibration accuracy twice that of the memory-less baseline.

The second longitudinal learning paper addressed a more fundamental question: do agentic systems with memory actually improve over time? The experiment design was clever. Treatment group: 3-line prompt + (S)AGE memory, able to query all accumulated knowledge from previous rounds each round. Control group: 50 to 200 lines of expertly crafted prompts, but no memory, starting from scratch each round. After 10 rounds, the treatment group's red-teaming difficulty score rose from 0.8 to 3.0 (Spearman rho=0.716, p=0.020), while the control group showed no growth trend (rho=0.040, p=0.901). The most critical point: there was no statistical difference in absolute performance between the two groups (Cohen's d = -0.07). A 3-line prompt with memory tied with a 200-line expert prompt. The difference lay in the learning trajectory: the system with memory got better over time, while the one without memory stayed at the same level.

This means the memory layer gives agentic systems not higher initial performance, but organizational-level longitudinal learning capability. A human organization's 100th project is usually better than its first because of process documentation, post-mortems, and accumulated knowledge bases. Now agentic systems are beginning to show the same pattern.

The differences among the three engineering paradigms can be summed up in one sentence: Prompt Engineering optimizes the interface between humans and models; Context Engineering optimizes the model's input space; Harness Engineering optimizes the entire runtime environment of Agents. Anthropic's experiments proved that evaluation loops are orders of magnitude more effective than self-evaluation. OpenAI's experiments proved that architectural constraints can maintain consistency in Agents at the million-line code level. The two papers proved that a consensus-verified memory system can give agentic organizations longitudinal learning capabilities. Together, these three layers form a complete Harness: evaluation mechanisms + architectural constraints + memory governance. Missing any one layer, and the agentic system will spin out of control in some dimension.

via Anthropic Engineering Blog / OpenAI Blog / (S)AGE Paper / Longitudinal Learning Paper