History of Development

1. Periodization Method and Reading Guide

图 1-1｜AI Harness 三代架构演进：L1–L6 层层叠加

数据来源：基于本文分析绘制的示意图。

1.1. Basis for Periodization

This document periodizes the evolution of the AI Harness using "the migration of the core bottleneck" as the basis, rather than a simple cut by year. The criteria are as follows:

1.2. Overview of the Three-Generation Division

This project uniformly uses the following three-generation division throughout; it must not be altered:

It should be noted that the three generations are not the later wiping out the earlier, but rather stacking layer upon layer. Prompt engineering is still necessary today, it is just no longer sufficient.

1.3. Timeline at a Glance

图 1-2｜时间轴速览：从奠基期到第三代

数据来源：基于本文分析绘制的示意图。

• 2022-10 ── ReAct / LangChain created
• 2023-02 ── Toolformer
• 2023-03~04 ── AutoGPT / BabyAGI
• 2023-06-13 ── ★ OpenAI Function Calling
• 2023-10-10 ── SWE-bench released
• 2024-08-13 ── SWE-bench Verified
• 2024-10-22 ── Anthropic computer use public beta
• 2024-11-25 ── ★ MCP released
• 2025-02-24 ── ★ Claude Code released
• 2025-03-11 ── ★ OpenAI Responses API + Agents SDK
• 2025-04-09 ── Google A2A + ADK
• 2025-04-16 ── ★ Codex CLI released
• 2025-06-23 ── A2A donated to Linux Foundation
• late 2025-06 ── "context engineering" naming chain
• 2025-10-16 ── Anthropic Agent Skills
• 2025-11-25 ── MCP 2025-11-25 specification
• 2025-12-09 ── ★ AAIF founded
• 2025-12-18 ── Agent Skills becomes an open standard
• 2026-02-05 ── Hashimoto "Engineer the Harness"
• 2026-02-11 ── ★ OpenAI "Harness engineering"
• 2026-03 ── A2A v1.0
• H1 2026 ── ★ GB/Z 185—2026 China agent-interconnection national standard
• 2026-07-28 ── MCP specification made stateless

2. Foundations: From ReAct to Function Calling (2022—2023.6)

2.1. Academic Predecessors: Teaching Models to Use Tools

Function Calling was productized in June 2023, but its academic lineage can be traced back at least to October 2022.

The significance of ReAct is that it was the first to formalize the "think—act—observe" loop, a loop that remains the backbone of every Agent Loop to this day.

2.2. The First Explosion of Autonomous Agents

In March—April 2023, AutoGPT and BabyAGI were released in succession (different sources give either 2023-03 or 2023-04 as the release month; this document uniformly states "March—April 2023"), triggering the first wave of "autonomous agent" hype.

Key events in the same period:

The value and the lessons of this wave are equally stark: it proved that "model + loop" can autonomously complete multi-step tasks, but it also exposed the uncontrollability of unconstrained loops — infinite loops, goal drift, and runaway costs became pervasive problems. These lessons directly gave rise to the orchestration and control requirements of the second generation.

2.3. Evaluation Becomes an Independent Stage

On 2023-10-10, SWE-bench was released. This was a turning point: for the first time, evaluation was separated from model-capability assessment and became a distinct industry stage.

The principle SWE-bench established is programmatic grading: if the test passes, it passes — there is no partial credit for "looks good but fails the test." This principle later became the methodological cornerstone of the entire Harness evaluation layer.

Its initial SOTA was Claude 2 + BM25 at 1.96% (Grade B, but this is the paper's own figure and is highly credible). The curve that started at 1.96% later became the primary yardstick for measuring Harness capability.

3. First Generation: The Prompt-Centric Era (c. 2020—2023.6)

3.1. Technical Characteristics

3.2. Representative Forms

• Early GPT-3 applications (2020—2022): mainly creative generation, text rewriting, and classification;
• Few-shot template engineering: guiding output format through carefully designed examples;
• AutoGPT prototypes: although they already had a loop, their "intelligence" came almost entirely from prompts, lacking tool boundaries and evaluation.

3.3. Reasons for Being Superseded

The first generation was not eliminated, but rather deemed insufficient. Three insurmountable flaws:

1. No tools: the model could only output text and could not act on the environment, so its capability boundary was locked in;
2. No state: no handoff across sessions, so long tasks inevitably get interrupted;
3. No evaluation: it could not answer "is this better or worse than last time," so engineering it was out of the question.

4. Second Generation: The Tool & Orchestration-Centric Era (2023.6—2025.2)

4.1. Starting Point: The Function Calling Release

On 2023-06-13, OpenAI released "Function calling and other API updates" (Grade A), formally starting the second generation.

The historical significance of Function Calling is that it turned "models calling external capabilities" from a prompt trick into a protocolized API contract. From then on, tools could be described structurally, reliably selected by models, and safely executed by systems.

4.2. The Boom and Divergence of Orchestration Frameworks

After Function Calling, orchestration frameworks flourished rapidly. The main representatives and their positioning:

The prevailing mindset of this stage was "Framework = solution." But by 2024—2025, the industry gradually realized that frameworks only provide building blocks and do not solve production problems — how to organize context, what counts as "done," and how to regress after a failure are all questions frameworks leave unanswered. This gap in expectations directly gave rise to the Harness concept.

4.3. Protocolization: The Birth of MCP

On 2024-11-25, Anthropic released "Introducing the Model Context Protocol" (Grade A). This was the most important event of the second generation and the main anchor of the third generation's standardization narrative.

MCP solves what Anthropic calls the "N×M" data integration problem: M model applications need to connect to N data sources. Without a standard that means N×M adapters; with a standard it becomes N+M.

Technically, MCP reuses the message-flow idea of the Language Server Protocol (LSP) (Grade B, Wikipedia). Before it, OpenAI's function-calling and the ChatGPT plugin framework were all vendor-proprietary approaches.

4.4. The Industrialization of Benchmarks

2024 was the year benchmarks were industrialized, and the first time the claim that "the harness determines the result" was quantified at scale.

The most noteworthy record of the second half of 2024 is: Claude 3.5 Sonnet + agentic harness reached 49.0%. It was the first time "Harness" was written into a public benchmark scorecard — the same model with different harnesses scoring differently became an industry consensus from then on.

4.5. Reasons for Being Superseded

The second generation's four unresolved flaws:

1. Tool integration is still N×M: although MCP exists, the ecosystem is not yet mature;
2. Context without governance: context rot in long tasks is pervasive, and no one systematically solves it;
3. Evaluation is still bolted on: benchmarks are external leaderboards, not a feedback loop built into the system;
4. Governance is manual: permission prompts interrupt frequently, being neither safe nor smooth.

These four flaws correspond exactly to the four cores of the third generation — context engineering, sandboxed execution, trajectory evaluation, and permission governance.

5. Third Generation: The Runtime & Evaluation-Centric Era (2025.2—present)

5.1. Starting Point: Claude Code and Terminal Agents

On 2025-02-24, Anthropic released Claude 3.7 Sonnet and Claude Code (Grade A). Claude Code was positioned as "our first agentic coding tool" and released in the form of a limited research preview.

Its official capability description: search and read code, edit files, write and run tests, commit and push to GitHub, and use command-line tools.

Subsequent milestones (Grade B, marked [to be verified]): on 2025-05-22 it went GA with the Claude 4 release; within seven weeks of launch it reached 350,000 DAU, over 1 million merged PRs, and roughly 1/4 of Anthropic's internal code commits were generated by it.

The significance of Claude Code is that it turned the harness from "a set of APIs in an SDK" into an "out-of-the-box product," letting users intuitively feel the value of the harness for the first time.

5.2. Official Built-Ins for Orchestration and Observability

On 2025-03-11, OpenAI released "New tools for building agents" (Grade A), with four major components:

Key signal: orchestration and observability are built in officially. Evaluation and observability (L5) are no longer bolted-on third-party tools but first-class citizens of a vendor SDK. The Assistants API is planned for sunset in mid-2026.

On 2025-04-16, OpenAI released Codex CLI (Grade A):

o3 and o4-mini were released the same day; two days earlier (04-14) the GPT-4.1 series was released, with 1M token context.

On 2025-05-16, Codex Cloud research preview was released; codex-1 is a SWE fine-tuned variant of o3, reaching 72.1% single-shot and 83.8% over eight attempts on SWE-bench Verified (Grade B, marked [to be verified]).

In the same period, Google released ADK at Google Cloud Next 2025 on 2025-04-09 (Apache 2.0, Python first, v1.0.0 on 2025-05-20), and released the A2A protocol the same day.

This trend of "officially built-in orchestration and observability" was extended in September 2026 from an SDK to a managed runtime: on 2026-09-10, OpenAI launched the execution-control layer behind Codex as a public beta of the Agents API (per the official Changelog, Grade A) — session management, context compression, failure recovery, and multi-agent coordination are hosted by OpenAI; the API itself carries no platform fee and is billed by model and tool usage, and the sandbox can be OpenAI-hosted, provided by one of nine partners (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel), or self-built infrastructure. The explicit pricing of the orchestration layer is thereby reduced to zero, with revenue deferred to model and sandbox consumption.

Concurrent developments on the coding-agent side (per the official Changelog, cross-checked in aggregate): between 2026-09-01 and 09-12, Claude Code shipped 12 consecutive releases (v2.1.257—270), of which 2.1.257 set Claude Fable 5.1 as the default model (1M token context, $10/$50 per million tokens, $0.25 per cached read), 2.1.260 added the /diff real-time diff panel, 2.1.261 added the /skill-doctor skill-usage diagnostics, and 2.1.269 added the claude plugin eval plugin-evaluation suite (sinking L5 evaluation capability into the plugin ecosystem); OpenAI Codex CLI 0.154.0 (09-09) added GPT-6 Astra to the model picker and landed on Amazon Bedrock; GitHub Copilot's 09-07 weekly update introduced the experimental adaptive model routing Project HydraFusion in the CLI; and Google released Gemini 3.8 Flash on 09-02 (per the official statement), pushing the "budget-tier models taking on agentic tasks via longer reasoning" route to about 73.8% on DeepSWE v1.1 (virtually on par with GPT-6 Astra's 74.1%).

The same official Changelog page also records an early-September 2026 expansion of runtime capabilities: the GPT-Live 1 full-duplex voice API went GA (09-10, $0.05/minute billed per second), GPT-6 Astra became generally available (09-03, with tool calling supported only via the Responses API, plus the introduction of asynchronous safety checks and mid-turn redirection), and Prompt Cache Diagnostics went GA (09-08); MCP published its roadmap on 08-22, listing agent message primitives, unification of HTTP-native transport, and agent identity (DPoP / workload identity federation) as priorities for the next specification cycle. Accompanying industry and security signals include: Cursor's release of Projects (Beta, 09-10, dispatching parallel sub-agents through a coordinator agent that writes no code — the L3 orchestration layer of the AI IDE begins to take object form); the security firm Accomplish's disclosure of Claude Code macOS sandbox-escape details (09-11, reported 2026-07-13, fixed in v2.1.247); and OpenAI's disclosure of internal data from its research organization — agent workdays to human workdays reaching a ratio of 3.1:1, with median researcher daily spend exceeding $600 (09-06, per the official statement) — so that "agent hours surpassing human hours" is for the first time supported by internal data from a leading laboratory.

5.3. Naming and Maturing of Context Engineering

The eleven days in late June 2025 were the process by which "context engineering" went from a grass-roots practice term to a common industry term (Grade B, dates cross-checked across multiple sources):

It should be added that the term has an earlier grass-roots practice origin: HumanLayer founder Dex Horthy (author of 12-Factor Agents) has been using it since August 2024, and his factor #3 is "own the context window." Horthy describes himself as "not an inventor, but giving a name to what peers were already doing" (Grade B).

Around 2025-09-29, Anthropic released "Effective context engineering for AI agents" (Grade A), establishing attention budget and context rot as first-class concepts, marking the layer's official methodological standing.

5.4. The Skills Mechanism and Progressive Disclosure

On 2025-10-16, Anthropic released Agent Skills (Grade A); on 2025-12-18 it became an open standard.

Anthropic's official claim is: "the amount of context that can be bundled into a skill is effectively unbounded." This shifts the L1 layer from "how to stuff more in" to "how to load on demand."

5.5. Standardization: The Founding of AAIF

On 2025-12-09, the Linux Foundation announced the formation of the Agentic AI Foundation (AAIF) (Grade A; GIGAZINE records 2025-12-10; this document uses 12-09).

Key ecosystem data at the time of founding (Grade A):

Two quotes worth recording:

"Donating MCP to the Linux Foundation as part of the AAIF ensures it stays open, neutral, and community-driven as it becomes critical infrastructure for AI." — Mike Krieger, Anthropic CPO

"We are seeing AI enter a new phase, as conversational systems shift to autonomous agents that can work together." — Jim Zemlin, Executive Director, Linux Foundation

The industrial tension in open governance materialized in 2026-09: on 2026-09-02 NVIDIA signed, and on 09-03 announced, an agreement of approximately $12.93 billion to acquire Hugging Face (approximately $11.9 billion to shareholders, plus up to $1 billion in employee retention equity; disclosed in an SEC 8-K, with closing expected in the first half of 2027, consistent across multiple sources), and the distribution layer of the open-model ecosystem thereby folded into a compute vendor. Official commitments include preserving the Hugging Face brand, multi-cloud and multi-accelerator support, and no mandatory use of NVIDIA compute, but the acquirer is simultaneously the largest open-source-model contributor on the platform (500+ models, 250+ datasets), and whether the "neutral hub" can be maintained after closing has become the industry variable most worth tracking outside the AAIF's neutral-governance narrative.

5.6. Naming Established: Harness Engineering

In February 2026, the term "harness" made the leap from industry jargon to an official term within six days.

Hashimoto's six stages are: Drop the Chatbot → Reproduce Your Own Work → End-of-Day Agents → Outsource the Slam Dunks → Engineer the Harness → Always Have an Agent Running.

In the same post, OpenAI disclosed its million-line code experiment (Grade A, all figures may be cited in full):

This experiment is the benchmark demonstration of the third-generation architecture; the details of its engineering decisions are in 03-架构演进.md.

5.7. China: The National Standard System for Agent Interconnection

In the first half of 2026, China released the GB/Z 185—2026 "Artificial Intelligence — Agent Interconnection" family of national standards (Grade A; there are three reported release dates — 2026-05-22 / 2026-06-26 / 2026-07-09 — and this document presents them side by side).

The seven-part structure:

1. Part 1: Overall architecture
2. Part 2: Identity codes (encoding, allocation, and management)
3. Part 3: Identity management (registration, accounts, credentials, authentication)
4. Part 4: Agent description (capability description and registration, publication, changes)
5. Part 5: Agent discovery (discovery process)
6. Part 6: Agent interaction (peer-to-peer, group, hybrid)
7. Part 7: External tool invocation (architecture, flow, data formats)

Deployment progress: on 2026-08-31, a special session on the application of the "Artificial Intelligence — Agent Interconnection" national standard (Yangtze River Delta region) was held in Zhangjiang, Pudong New Area, Shanghai, where 30 organizations including AISpeech, Lenovo, ZTE, Volcano Engine, and Shanghai Jiao Tong University were awarded "agent identity code node" status (Grade A).

For the standard's details and trend assessment, see Section 3.5 of 04-未来发展.md.

6. Master Table of Key Milestones

The table below summarizes all the key milestones covered in this document, sorted by time, with the source grade of each marked.

7. Comparison of the Three Generations

8. Evolutionary Patterns and Implications

From the evolution across the three generations, four regularities can be distilled:

Rule 1: The bottleneck keeps moving outward.

From "the model can't use tools" (1st generation) to "too many tools to connect" (2nd generation) to "context and governance determine success or failure" (3rd generation), each generation solves the external problems exposed by the previous one, not the problems internal to the model.

Rule 2: Each generation sinks the core means of the previous one into infrastructure.

Prompt engineering was not wiped out, but sank into part of the L1 layer; tool invocation was not wiped out, but was standardized as MCP. A new generation does not replace — it turns the previous generation's manual work into default capabilities.

Rule 3: Standardization always follows the emergence of a de facto standard.

MCP was released in 2024-11 but only donated to the Linux Foundation in 2025-12; AGENTS.md was adopted by 60,000+ projects before becoming an AAIF donation; Agent Skills was released in 2025-10 but only became an open standard in 2025-12. The regularity is: first a widely adopted de facto standard, then a legally standardized one under neutral governance.

Rule 4: Naming lags practice by about 12—18 months.

The practice of context engineering began in August 2024 (12-Factor Agents), yet was only named in June 2025; the product practice of the harness began in early 2025 (Claude Code / Codex CLI), yet was only named in February 2026. This suggests: by the time a concept is formally named, its practice has usually been running for at least a year.

9. Summary

The history of the AI Harness is essentially a history of "the migration of engineering focus from inside the model to outside the model".

• The 1st generation focused on how to say it (Prompt);
• The 2nd generation focused on what it can do (Tools + Orchestration);
• The 3rd generation focused on in what environment it acts, whether it does it well, and whether it can be constrained (Context + Sandbox + Eval + Governance).

Three key dates frame this migration: on 2023-06-13 Function Calling gave tools a contract; on 2024-11-25 MCP gave the tool ecosystem a standard; and on 2025-12-09 AAIF gave the standard neutral governance. Then the release of OpenAI's "Harness engineering" on 2026-02-11 marked the moment this body of engineering practice finally got a name of its own.

The stage the industry is currently (2026) at is: the third-generation architecture is established, the standardization window is closing, but the methodology has not yet converged. For the direction of the next phase, see 04-未来发展.md.

10. Information-Gap Statement

This document involves the following information gaps, marked at the corresponding locations in the body text:

1. ISO/IEC international standard for agent interconnection: no ISO/IEC number for a published or initiated international standard on agent interconnection was found; no authoritative information is currently available.
2. The first use of the term "Agent Harness": no exact primary source was found; no authoritative information is currently available.
3. The LangChain "Anatomy of an Agent Harness" original text and date: the original text was not found; no authoritative information is currently available.
4. The exact URL and release date of Anthropic's "How we built our multi-agent research system": it is only referenced as a link in the context-engineering blog and has not been verified.
5. Conflicting release dates for GB/Z 185—2026: 2026-05-22 (China Daily) / 2026-06-26 (Baidu Baike) / 2026-07-09 (People's Daily report date, with the wording "recently released"). This document presents them side by side; the body uses "first half of 2026."
6. Conflicting AAIF founding dates: 2025-12-09 (TechCrunch, Wikipedia) / 2025-12-10 (GIGAZINE). This document uses 12-09.
7. Conflicting Terminal-Bench 2.0 release timing: one source records late 2025, another 2026-01. This document writes "late 2025 to early 2026."
8. Conflicting Anthropic computer use release dates: Taskade records 2023-10, conflicting with the official 2024-10-22. This document follows the official 2024-10, marked [uncertain].
9. Conflicting AutoGPT / BabyAGI release months: two accounts, 2023-03 and 2023-04. This document writes "March—April 2023."
10. The "first public spec version" wording of MCP: 2024-11-05 (the Ruby SDK records it as the Initial protocol revision) and 2024-11-25 (public announcement and ecosystem launch) can coexist.
11. All Grade B/C figures in this document are marked [to be verified], including: the Devin release date; SWE-bench Verified's 500/93/68.3%/33.2%; the various scores of the second half of 2024; Claude Code's 350,000 DAU / 1 million PRs / 25% internal commits; codex-1's 72.1% / 83.8%; and the Claude 2025-11-24 version date.
12. The 2026 version timeline of Claude models: found only in a third-party GitHub timeline repository; it needs to be checked against official announcements and is marked [to be verified].

11. References

1. Function calling and other API updates — OpenAI, 2023-06-13. https://openai.com/blog/function-calling-and-other-API-updates
2. Introducing the Model Context Protocol — Anthropic, 2024-11-25. https://www.anthropic.com/news/model-context-protocol
3. Claude 3.7 Sonnet and Claude Code — Anthropic, 2025-02-24. https://www.anthropic.com/news/claude-3-7-sonnet
4. New tools for building agents — OpenAI, 2025-03-11. https://openai.com/blog/new-tools-for-building-agents
5. Introducing Codex (Codex CLI) — OpenAI, 2025-04-16. https://github.com/openai/codex
6. Agent Development Kit: Making it easy to build multi-agent applications — Google, 2025-04-09. https://googledevelopers.blogspot.com/en/agent-development-kit-easy-to-build-multi-agent-applications/
7. A year of open collaboration: Celebrating the anniversary of A2A — Google Open Source Blog, 2026-04-16. https://opensource.googleblog.com/
8. Harness engineering: leveraging Codex in an agent-first world — OpenAI, 2026-02-11. https://openai.com/index/harness-engineering/
9. My AI Adoption Journey — Mitchell Hashimoto, 2026-02-05. https://mitchellh.com/writing/my-ai-adoption-journey
10. Harness engineering for coding agent users — Birgitta Böckeler, martinfowler.com, 2026. https://martinfowler.com/articles/harness-engineering.html
11. Effective context engineering for AI agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
12. Introducing Agent Skills — Anthropic, 2025-10-16 (open-standard update 2025-12-18). https://www.anthropic.com/news/skills
13. Linux Foundation Announces the Formation of the AAIF — Linux Foundation / AAIF, 2025-12-09. https://aaif.io/
14. Linux Foundation Launches the Agent2Agent Protocol Project — Linux Foundation, 2025-06-23. https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents
15. MCP Protocol Versions — MCP Ruby SDK, 2026. https://ruby.sdk.modelcontextprotocol.io/protocol-versions/
16. ReAct: Synergizing Reasoning and Acting in Language Models — Yao et al., 2022-10 (arXiv:2210.03629). https://arxiv.org/abs/2210.03629
17. Toolformer: Language Models Can Teach Themselves to Use Tools — Schick et al. (Meta AI), 2023-02 (arXiv:2302.04761). https://arxiv.org/abs/2302.04761
18. SWE-bench: Can Language Models Resolve Real-World GitHub Issues? — Jimenez, Yang, et al. (Princeton), 2023-10. https://www.swebench.com/
19. "Artificial Intelligence — Agent Interconnection" family of national standards (release report) — People's Daily Online, 2026-07-09. https://finance-app.people.cn/n1/2026/0709/c1004-40757059.html
20. Interpretation of the "Artificial Intelligence — Agent Interconnection" family of national standards — China Industrial Economy Information Network, 2026. https://cinic.org.cn/xw/zcdt/1643418.html