Future Development

1. Approach and Reading Guide

图 1-1｜AI Harness 未来发展：三条主线、因果链与 12—24 个月展望

数据来源：基于本文分析绘制的示意图。

1.1. Three Trend Lines

This document decomposes the future development of AI Harness into three relatively independent but mutually influential main lines:

A causal chain runs among the three main lines: technology determines possibility, standards determine interoperability, and industry determines the intensity of investment.

1.2. Factual Baseline of This Document

Because this document makes judgments about the future, it must strictly distinguish facts from speculation:

• Anything sourced at grade A is an official first-hand statement from a vendor or organization and may be quoted directly;
• For anything sourced at grade B/C, its figures are always marked [To be verified];
• All inferences and judgments given by this document uniformly begin with "Judgment:", kept clearly separate from factual statements;
• For matters on which this research obtained no authoritative information, this document honestly states "no authoritative information" rather than making speculative filler.

---

2. Technology Trends

2.1. Co-evolution of the Model and Harness

This is the most important — and most tension-laden — of the technology trends.

Trend direction (grade A, Anthropic's official self-description): As more capable models are released one after another, the models themselves can already cover some capabilities that previously required Harness to fill in. Anthropic states plainly that the token cost of context reset and the coordination cost of the Sprint structure have actually become new friction on stronger models. The company says it is "dynamically removing redundant Harness design as model capability improves".

Counter-trend (grade A, same source): Products such as Claude Code / Codex are already training the model and Harness together during the post-training stage, producing the side effect of the "model being overfit to a specific Harness structure" — changing the tooling logic actually makes the model perform worse.

The practical implication of this two-way effect is:

The best Harness for your task is not necessarily the one the model was trained with during post-training.

Judgment: The relationship between the model and Harness is not one-way replacement but a reallocation of responsibilities during co-evolution. As models grow stronger they absorb part of Harness's capabilities (for example, longer effective context reduces reliance on complex compression strategies), yet at the same time they add new Harness requirements (for example, a model that can run tasks longer than 6 hours needs stronger interruption-recovery and cost-control mechanisms). So the total amount of Harness will not go to zero, but its composition will keep changing.

2.2. The Agent Version of the Bitter Lesson: Architecture Lightweighting

Facts (grade B/C; case figures marked [To be verified]):

The conclusion the community draws from this is:

Every time a new model is released, the best agent structure changes — what required complex hand-built pipelines in 2024, a single context window's prompt will suffice for in 2026.

This judgment is Rich Sutton's The Bitter Lesson, in the agent domain: general methods that exploit computation and scale ultimately prevail over clever designs that rely on human priors.

Operational implication: Harness must be lightweight, ready at any moment to discard yesterday's "clever" logic. Concretely:

1. Avoid hard-coding business rules into orchestration graphs;
2. Prefer general structures the model can understand directly (filesystem, git, standard CLI);
3. Annotate each Harness component with the condition for "when it can be deleted".

2.3. Harness as a Dataset: the Shift of Competitive Advantage

Facts (grade B, relayed from industry surveys): In the future Harness will become the main tool for addressing "model drift" — labs use Harness to precisely detect when a model stops following instructions or makes reasoning errors after step 100, and these data feed directly back into training.

Core claim:

The trajectories your Harness captures are where the competitive advantage lies, no longer the prompts themselves.

Open frontiers being explored (grade B):

1. Running hundreds of agents in parallel on a shared codebase;
2. Having agents analyze their own trace logs to identify and fix Harness-level faults;
3. Harness being able to dynamically assemble exactly the right tools and context, rather than pre-configuring everything.

Judgment: If this trend holds, Harness will shift from a "cost center" to a "data asset". An enterprise's Harness accumulation (trajectories, evaluation sets, failure-mode libraries) will form a moat just like training data, and this asset is not attainable by purchasing models.

2.4. Deepening of Context Engineering

From the naming chain of 2025-06 to 2026, context engineering has moved from "concept" to "an engineering discipline with an official methodology". Three directions for further deepening:

At the same time, the four pillars proposed by Lance Martin (Write / Select / Compress / Isolate) and Anthropic's four strategies (compaction / note-taking / memory tool / sub-agent) have formed two mutually translatable taxonomies, showing that this layer is converging into a standard terminology system.

2.5. Industrialization of Evaluation and Observability

Facts (grade A): Evaluation has moved from external leaderboards into being internalized as a system component. Three signs:

1. OpenAI builds tracing into the Agents SDK;
2. Anthropic builds a standalone Evaluator with hard-threshold scoring into long-running Harness;
3. Google ADK includes AgentEvaluator.evaluate(), which can score both the final response and the step-by-step trajectory.

Open problems that have emerged (grade A, from martinfowler.com):

We need Harness coverage and quality evaluation methods analogous to code coverage and mutation testing.

Two more evaluation-design ideas worth recording:

Judgment: The evaluation layer is the least mature, yet highest-marginal-return, layer of the third-generation architecture. The breakthrough most likely in the next 12–24 months is not the model but evaluation methodology — especially attempts like "Harness coverage" that transplant software-engineering maturity models.

2.6. The Runtime Sinking into System-Level Languages

Facts (grade B, marked [To be verified]): Codex CLI was rewritten from Node.js to Rust in the second half of 2025; by early 2026 roughly 95% of the codebase is Rust. The official reasons: static single-file distribution, no GC, low memory footprint, and direct access to OS sandbox primitives.

Judgment: This is a signal that Harness is moving from "application-layer tool" toward "system software". When Harness needs to run for long periods (single tasks over 6 hours), needs fine-grained control over sandbox primitives, and needs to reside quietly on a user's machine, the system-level characteristics of the runtime language shift from a "nice-to-have" to a "requirement".

---

3. Standardization Trends

3.1. Global Overview: a Layered Standards Stack Is Taking Shape

The current agent standards stack already shows clear layering (grade A, synthesized from the AAIF website, Google Open Source Blog, People's Daily, and others):

The A2A family has further extended into AP2 (Agent Payment Protocol), A2UI (Agent to UI), and UCP (Universal Commerce Protocol).

3.2. Neutral Governance: the Tension in AAIF and Open Governance

The Agentic AI Foundation (AAIF), established on 2025-12-09, marks standardization entering a phase of neutral governance.

But tension exists within open governance, and two quotes are worth placing side by side:

OpenAI engineer Nick Cooper voiced concern about standards ossification:

"I don't want it to be a stagnant thing. I don't want these protocols to be part of this foundation, and that's where they sat for two years. They should evolve."

Linux Foundation executive director Jim Zemlin, meanwhile, acknowledged the risk that "de facto standards overwhelm de jure standards" — even with open governance, one company's implementation can still become the default simply by being released fastest and used most, drawing the Kubernetes analogy that "dominance emerges from merit and not vendor control".

Judgment: Over the next 12–24 months, the biggest thing to watch in standardization is not "whether there will be a standard" (there already is), but "whether the standard will evolve or freeze". Whether AAIF's governance mechanism can keep pace with the iteration speed of models and Harness will determine whether MCP becomes the next Kubernetes or the next CORBA.

3.3. Protocol Layer: MCP's Statelessness and Extension Framework

The 2026-07-28 version of the MCP specification is the most important recent change at the protocol layer (grade A):

The community's judgment on this direction is: "statelessness is the direction of travel" (grade B).

Judgment: Statelessness is the necessary path for the protocol to move from "developer tool" to "cloud-native infrastructure". This change sits in potential tension with Harness's long-running needs (single tasks over 6 hours) — long tasks inherently require state tracking, which is precisely the Tasks capability externalized as an Extension. A stateless core plus stateful extensions will be the typical shape of future protocol design.

3.4. Multi-Agent Interoperability: A2A and the Protocol Family

Google's official definition of the division of labor between MCP and A2A (grade A):

"MCP manages internal tool integration, while A2A handles the vital external coordination between autonomous entities."

Judgment: The release of A2A v1.0 and Signed Agent Cards means multi-agent interoperability moves from the "can it connect" stage to the "can it be trusted" stage. Signed Agent Cards address "how to confirm that another agent's identity-and-capability claims have not been tampered with", which is two solution paths to the same problem as China's GB/Z 185 series' "identity code + identity management".

3.5. China: the GB/Z 185—2026 Agent Interconnection Standard System

This is the finding with the greatest unique value in this research: China has released the world's first systematic national standard system for agent interconnection.

3.5.1. Basic Information on the Standard

3.5.2. Seven-Part Structure

These seven parts form a complete closed loop: identity code (who you are) → identity management (how to prove it) → description (what you can do) → discovery (how to find it) → interaction (how to collaborate) → tool invocation (how to act).

3.5.3. Official Definition of an Agent

The official Chinese definition of "agent" can serve as the standard wording in the Chinese context (grade A, standard interpretation from China's Industry Economy Information Network):

"An agent is a new type of intelligent system equipped with autonomous perception, memory, decision-making, interaction, and execution capabilities, and has become the key carrier of AI products and services."

3.5.4. Implementation Progress

3.5.5. Follow-on Plans

The follow-on actions the authorities have already made explicit (grade A):

1. Promote, at an appropriate time, the conversion of identity-code-related standards into mandatory national standards;
2. Accelerate the development of substandard such as agent auditing and agent trading.

Item 2 is worth noting: the two directions of "agent auditing" and "agent trading" correspond exactly to the L6 governance and security layer of this project's six-layer model and to the commercial closed loop of industrialization — indicating that China's standard system is extending from "interconnection" toward "auditable and tradable".

3.5.6. The Broader Chinese Standards Context

3.5.7. Judgments

Judgment 1 · The divergence of the two routes stems from different roots.

The AAIF route is "bottom-up, retroactive recognition of de facto standards" — first MCP is widely adopted, then donated to the foundation; the GB/Z 185 route is "top-down, leading by front-loaded top-level design" — laying out a complete framework of identity, description, discovery, interaction, and tool invocation before the industry has settled. The former's strength is ecosystem breadth; the latter's strength is closed-loop completeness.

Judgment 2 · Identity is the core lever of the Chinese route.

Of GB/Z 185's seven parts, two (Parts 2 and 3) concern identity directly, and they explicitly state that "one identity code can correspond to only one agent" and that "two-way identity authentication is required before interaction", with a plan already in place to convert to a mandatory national standard. This sends a clear signal: agent identity will be the entry point of Chinese regulation. For Harness products operating in China, identity-code compatibility is a compliance item that needs to be planned in advance.

Judgment 3 · The "guidance technical document" form of GB/Z is worth watching.

Choosing GB/Z rather than GB/T or a mandatory GB indicates the authorities' judgment that "the industry is still changing rapidly, so it should not be locked down now". This is a pragmatic agile standardization arrangement. But the phrasing "promote, at an appropriate time, the conversion of identity-code-related standards into mandatory national standards" simultaneously implies: the window period is limited.

3.6. Comparing the Two US-China Standardization Routes

Judgment: The two routes will not conflict in the short term, because they cover different layers — AAIF focuses on "how to connect", while GB/Z focuses on "who is connecting". But as both A2A's Signed Agent Cards and GB/Z's identity codes enter implementation, mutual recognition of identity systems will become a key issue over the next 24 months.

---

4. Industry Trends

4.1. Adoption Has Saturated, but Trust Has Not Caught Up

The complete picture from Stack Overflow's 2025 annual developer survey (grade A, released 2025-07-29, 49,000+ responses, 177 countries, 62 questions, 314 technologies):

Reasons for turning to a human instead of AI: 75.3% don't trust AI answers, 61.7% have ethical or safety concerns, 61.3% want to fully understand their own code.

Judgment: This data set delineates Harness's market space — a 84% adoption rate means the market no longer needs education, while the 46% distrust rate and the 3.1% high-trust rate mean predictability is still the core unmet need. And predictability is precisely Harness's defining responsibility. The larger the trust gap, the larger Harness's market space.

4.2. Enterprise-Side Efficiency Data

The following data is relayed from aggregator sites; all of it is marked [To be verified]; it is advised to cross-check the primary reports before use:

Judgment: DORA's "negative correlation with delivery stability" is the item most worth heeding. It shows that current AI adoption mainly boosts speed rather than quality — which is exactly the problem Harness (especially L5 evaluation and L6 governance) should solve. Conversely, this also explains why Harness became a standalone concept in 2025–2026: once the speed dividend is exhausted, engineering focus necessarily shifts to stability.

4.3. Market and Industry Movements (Including the Chinese Market)

The following items are all marked [To be verified]; they are only seen as relays from encyclopedia-style entries and should be double-checked before being used for decisions:

Judgment: If these three movements are accurate, they point to a common pattern — model vendors are extending into the Harness layer. This is consistent with the paths of Anthropic and OpenAI: the boundary between model companies and Harness companies is blurring, and an integrated "model + dedicated Harness" delivery may become the mainstream form. At the same time, the direction of "industrial embodied Harness" suggests the Harness concept is spilling beyond software engineering into embodied intelligence and industrial scenarios.

Concurrent industry events add two further lines of evidence from the supply side. Managed runtimes and consolidation of the compute–distribution layer: the OpenAI Agents API public beta (2026-09-10, per the official statement) prices the orchestration layer explicitly at zero, deferring revenue to model and sandbox usage; NVIDIA signed on 09-02 and announced on 09-03 an agreement to acquire Hugging Face for approximately $12.93 billion (SEC 8-K, expected to close in the first half of 2027), folding the distribution layer of the open-model ecosystem into a compute vendor; Figure AI reached a compute partnership with Nscale (initial commitment of $3.5 billion, planning for up to 100,000 Vera Rubin GPUs, per an industry weekly report, no official announcement seen, [To be verified]). Agenticization of economy-tier models: Google released Gemini 3.8 Flash on 09-02 (per the official statement), pushing DeepSWE v1.1 to about 73.8% at a unit price on par with the previous generation ($0.75/$3.75 per million tokens), while multiple evaluations report higher token consumption (under some evaluation calibers, per-task cost rises by about 40%) — cost measurement is shifting from "unit price" to "task cost", further raising the value of L5 cost observability.

Developments on the Chinese side corroborate the spillover of the Harness concept into embodied intelligence: Kinetix AI (Chaowei Dongli) completed a series of angel-plus rounds exceeding 500 million yuan on 2026-09-11 (led by Fountain Venture, with Far Light Ventures and Wanshi Capital following, confirmed by multiple authoritative media sources); XPeng's robotics business completed its first funding round of over $900 million on 2026-08-24 (post-money valuation of about $6.3 billion, a domestic record for a single embodied-intelligence round; consistent across multiple media sources, no official announcement seen); Guangzhou Yunding Technology unveiled RoboForge, a system-level embodied brain, at the World Robot Conference and XAIR Expo (09-03–06) — placing a system layer between the foundation model and robot execution, with a built-in Harness module that records the full-process evidence chain of tasks and converts failure samples into iteration material (jointly developed with Nanyang Technological University, Singapore; its top-ranking results on Embodied Bench and LaRA are vendor self-reported figures). "System-level embodied Harness" now has the dual support of named products and capital supply.

4.4. Market-Size Forecasts

The following data are marked [To be verified], and the measurement scopes differ enormously:

Judgment: The two figures differ by nearly an order of magnitude, rooted in different market scopes (models / full stack / AI-enabled services). This project recommends citing them only directionally, not as decision inputs. No authoritative estimate for the market size of the Harness layer itself was found in this research; it is marked [To be filled].

---

5. Outlook for the Next 12 to 24 Months

Synthesizing the three main lines above, the following outlook can be offered (the following are judgments, not statements of fact):

Two uncertainties need to be emphasized:

1. The cadence of model capability leaps is unpredictable. If a capability step-change occurs, most of the architectural judgments in this document would need to be re-evaluated within 6 months.
2. This research found no agent-interconnection international standard at the ISO/IEC level; there is currently no authoritative information. If progress occurs at that level, the judgment of the global standardization landscape would need to be rewritten.

---

6. Summary

The future development of AI Harness can be summed up in three sentences:

Technically, it will get lighter but will not disappear. The model and Harness are in a co-evolutionary relationship, not a substitute one. Stronger models absorb part of Harness's responsibilities while also adding another part; and the post-training coupling of the model to a specific Harness raises the importance of Harness design. The agent version of the bitter lesson reminds us: Harness must stay lightweight and be ready at any moment to discard yesterday's "clever" logic.

On standards, the window is closing. AAIF has established neutral governance, MCP has established a deprecation policy and evolution mechanism, and A2A has released v1.0 with Signed Agent Cards. China, meanwhile, has used the seven parts of GB/Z 185—2026 to build the world's first systematic agent-interconnection standard system, with the identity code as the core lever and a plan to convert to mandatory national standards. The two US-China routes are complementary in the short term and will need mutual recognition at the identity level in the long term.

On industry, adoption has saturated while trust has not caught up. The huge gap between the 84% adoption rate and the 3.1% high-trust rate is precisely Harness's market space; the "throughput up, stability down" revealed by DORA is precisely the problem Harness should solve.

One-Sentence Conclusion: AI Harness is moving from an "engineering technique" to an "engineering discipline", from a "cost center" to a "data asset", and from "vendors' private property" to "public infrastructure".

---

7. Information-Gap Declaration

This document involves the following information gaps, already marked in the corresponding places in the main text:

1. Agent-interconnection international standard at the ISO/IEC level: No ISO/IEC standard number for agent interconnection was found, either published or under development; there is currently no authoritative information. This directly affects the completeness of the global standards stack in Section 3.1.
2. Conflict over the GB/Z 185—2026 release date: Three datelines coexist — 2026-05-22 / 2026-06-26 / 2026-07-09. This document presents them side by side; the main text uses "the first half of 2026".
3. Current data from the Terminal-Bench official site and leaderboard: Not directly scraped; no authoritative first-hand data currently.
4. Official DORA 2025 URL: Only a relay of https://dora.dev/dora-report-2025 is seen, and its accessibility is unverified; all DORA figures in Section 4.2 are marked [To be verified].
5. Official URL of McKinsey's The State of AI 2025: Same as above; all McKinsey figures in Section 4.2 are marked [To be verified].
6. All JetBrains 2025 figures: Relayed from aggregator sites, marked [To be verified].
7. Chinese vendor movements (DeepSeek 2026-05-20, Xiaomi 2026-06-11, Lingxi Zhiyong 2026-08): All seen only as relays from encyclopedia-style entries, marked [To be verified]; it is strongly advised to double-check them before using them for decisions.
8. Manus's five re-architectures, LangChain's three redesigns, Vercel's removal of 80% of tools: Grade B/C sources; the case figures are marked [To be verified].
9. Codex CLI's Rust rewrite ratio (roughly 95% by early 2026): Grade B source, marked [To be verified].
10. Timeline of Claude model versions in 2026: Seen only in a third-party GitHub timeline repository, marked [To be verified].
11. Market-size forecast scope differences: Grand View Research and Bloomberg Intelligence differ by nearly an order of magnitude; no authoritative estimate for the market size of the Harness layer itself was found in this research; it is marked [To be filled].
12. The AGNTCY project: Seen only in AAIF-related survey mentions, not deeply verified; no authoritative first-hand material currently.
13. Directions recommended for additional searching: group standards / industry specifications from CAICT and AIIA; security-side specifications such as the OWASP Agentic AI Top 10; IEEE standard projects on AI agents; clauses in the EU AI Act related to autonomous agents. None of these four were searched this time.

---

8. References

1. Harness design for long-running application development — Anthropic, 2026. https://www.anthropic.com/engineering/harness-design-long-running-apps
2. Effective context engineering for AI agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
3. Effective harnesses for long-running agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
4. Harness engineering for coding agent users — Birgitta Böckeler, martinfowler.com, 2026. https://martinfowler.com/articles/harness-engineering.html
5. Humans and Agents in Software Engineering Loops — martinfowler.com, 2026. https://martinfowler.com/articles/exploring-gen-ai/humans-and-agents.html
6. Harness engineering: leveraging Codex in an agent-first world — OpenAI, 2026-02-11. https://openai.com/index/harness-engineering/
7. New tools for building agents — OpenAI, 2025-03-11. https://openai.com/blog/new-tools-for-building-agents
8. Linux Foundation Announces the Formation of the AAIF — Linux Foundation / AAIF, 2025-12-09. https://aaif.io/
9. OpenAI co-founds the Agentic AI Foundation under the Linux Foundation — OpenAI, 2025-12. https://openai.com/index/agentic-ai-foundation/
10. Donating the Model Context Protocol and establishing the AAIF — Anthropic, 2025-12. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
11. MCP Protocol Versions — MCP Ruby SDK, 2026. https://ruby.sdk.modelcontextprotocol.io/protocol-versions/
12. Model Context Protocol official site and specification — MCP / AAIF. https://modelcontextprotocol.io/specification/2026-07-28/
13. A year of open collaboration: Celebrating the anniversary of A2A — Google Open Source Blog, 2026-04-16. https://opensource.googleblog.com/
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. Release report on the Artificial Intelligence — Agent Interconnection series of national standards — People's Daily, 2026-07-09. https://finance-app.people.cn/n1/2026/0709/c1004-40757059.html
16. Interpretation of the Artificial Intelligence — Agent Interconnection series of national standards — China Industry Economy Information Network, 2026. https://cinic.org.cn/xw/zcdt/1643418.html
17. GB/Z 185—2026 implementation: first batch of Yangtze River Delta agent identity-code nodes issued — China Daily, 2026-09-04. https://cn.chinadaily.com.cn/a/202609/04/WS6a9a6773e4b09a165c788098.html
18. 2025 Stack Overflow Developer Survey — Stack Overflow, 2025-07-29. https://survey.stackoverflow.co/2025/
19. Stack Overflow 2025 Developer Survey official press release — Stack Overflow, 2025. https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/
20. AI Adoption Stats & Trends (DORA / McKinsey / JetBrains aggregation) — daily.dev, 2026-07-19. https://daily.dev/agentic-ai-hub/ai-adoption-stats-trends/
21. Agent Harness: the core paradigm of AI engineering in 2026 — Tencent Cloud Developer Community, 2026. https://developer.cloud.tencent.com/article/2698416
22. Harness architecture — Baidu Baike, 2026. https://baike.baidu.com/item/Harness%E6%9E%B6%E6%9E%84/67704948
23. Model Context Protocol — Wikipedia. https://en.wikipedia.org/wiki/Model_Context_Protocol
24. Agentic AI Foundation — AI Wiki. https://aiwiki.ai/wiki/agentic_ai_foundation
25. 12-Factor Agents — Dex Horthy / HumanLayer (URL). https://github.com/humanlayer/12-factor-agents