TL;DR
From August 14-25, 2026, the AI Agent track witnessed 6 major Harness Engineering events erupt in 12 days: ① 8/14 DeepSeek Harness (DSH) open-sourced under MIT license — black-whale WeChat branding, 173,700 GitHub Stars + 17,880 Forks in 5 days, HN 744 points 310 comments, “Everything is a plugin” philosophy, Preview stage; ② 8/19 OpenAI Codex Harness fully open-sourced under Apache-2.0 (github.com/openai/codex) — 107,443 Stars + 16,354 Forks, v0.149.0 stable released 8/20, 3-tier integration interfaces (light non-interactive codex exec + programmatic Codex SDK + persistent-session Codex app-server), covering all scenarios from CI scripts to product-grade Agents; ③ Microsoft Agent Lightning v1.0 + v1.0.1 (8/24-25 GA) — 17,600 Stars, 3,500-line core code, zero-code-change Agent RL training, 6,000 training samples lift Qwen3.5-9B SWE-bench Verified from 41.8% to 56.4% (+14.6), native Kubernetes Job + Agent Lightning Skill (lets Claude Code/Codex/Copilot systematically optimize other AI Agents’ prompts/tools/workflows); ④ Omarchy 4.0 “AI Operating System” — DHH + Dell + Jack Dorsey + 5 co-founders 8-person Omacom Foundation, 8M USD first-round, Linux distro with AI coding Agent built into the system, 10 days 1,000+ community plugin contributions, “new OS layer above models, below hardware”; ⑤ xAI Grok Bot 8/22-23 launches 24/7 autonomous cloud Agent — independent cloud computer autonomously managing tool flows 24/7, integrated into SuperGrok Heavy + Cursor Ultra + Cursor Teams; ⑥ ByteDance 8/24 merges TRAE + Coze office lines into Doubao (Zhao Qi unified reporting), Feishu aily renamed “Doubao Work Partner”; in same period Hugging Face explores 13B USD sale. Agent paradigm shifts from “standardization protocol layer” (MCP/Agent Plugins) down to “Harness execution layer” — “Agent = Model + Harness” becomes industry consensus: large models are engines, Harness is the engineering system that turns the engine into a controllable car that can actually run on the road.
I. Today Headlines: 8/14-8/25 Agent Track 6-Event Eruption in 12 Days “Harness Engineering”
| Dimension | 8/14-8/25 Key Event | Source | Strategic Significance |
|---|---|---|---|
| DeepSeek Harness (DSH) | 8/14 MIT open-source, 5 days 173.7K Stars + 17.88K Forks, HN 744 points | Tencent News 8/22 + Studio Global AI 8/24 + AI Disruption 8/24 | ”Everything is a plugin” + separate black-whale WeChat |
| OpenAI Codex Harness | 8/19 Apache-2.0 fully open-sourced, 107,443 Stars + 16,354 Forks, 3-tier integration interfaces | Qiniu Cloud 8/21 + Weibo 8/21 + OpenAI official + Tencent News 8/22 | ”Codex as a platform” — escalating from product to platform |
| Microsoft Agent Lightning v1.0 | 8/24-25 GA + v1.0.1 Skill, 17.6K Stars, 6000-sample Qwen3.5-9B SWE-bench +14.6 (41.8%→56.4%) | cnblogs 8/25 + GitHub microsoft/agent-lightning | ”Use production-time real Agent Harness for RL training” |
| Omarchy 4.0 AI OS | 8/20+ DHH + Dell + Dorsey 8-person Omacom Foundation 8M USD, 10 days 1,000+ plugins | Toutiao 8/25 + Tencent Securities 8/25 | ”Above models, below hardware” new OS-layer opportunity |
| xAI Grok Bot | 8/22-23 24/7 autonomous cloud Agent, integrated into SuperGrok Heavy + Cursor Ultra + Cursor Teams | HeadsUpAI 8/22 + HeadsUpAI 8/23 | Always-on autonomous agent teammate paradigm |
| ByteDance TRAE + Coze into Doubao | 8/23-24 office lines merged, Zhao Qi unified reporting, Feishu aily renamed “Doubao Work Partner” | Toutiao 8/25 + so.html5.qq.com 8/25 | First time a Chinese internet giant fully integrates Agent product lines |
| Hugging Face exploring 13B USD sale | 8/23 hired bankers approaching buyers, valuation 4.5B (2023) → 13B (+189%) | Tencent News 8/24 + Tencent Securities 8/25 | AI middle layer being bought up + open-source community neutrality concerns |
Core Judgment: 8/14-25 marks the inflection point where AI Agent shifts from “standardization protocol layer” (MCP/Agent Plugins) to “Harness engineering execution layer” — completely different from the historical 6 P5 angles (8/15 enterprise FDE rollout-depth / 8/7 standardization protocol year-one / 7/21 WAIC consumer explosion / 7/13 State Council 3-year plan / 6/29 scissors gap / 6/7 7 major trends). The shift is from “protocol layer (MCP/Anthropic SDK)” to “Harness layer (Codex Harness/DSH/Agent Lightning/Omarchy/Grok Bot)” + “vendor consolidation layer (ByteDance Doubao merger)” + “open-source platform acquisition layer (Hugging Face 13B)” three-axis resonance.
II. Bullet 1: 8/14 DeepSeek Harness (DSH) MIT Open-Source — 5 Days 173.7K Stars + HN 744 Points + “Everything is a Plugin”
2.1 Key Hard Data
Source: Tencent News 8/22 “After Black Whale, OpenAI Opens Harness, Competing for ‘Agent Runtime’” (news.qq.com/rain/a/20260822A027E700) + Studio Global AI 8/24 “DeepSeek Harness v0.1: The Open-Source Agent Framework That Broke GitHub Records” (studioglobal.ai) + AI Disruption 8/24 “DeepSeek Harness vs. Codex Harness: A Clash of Engineering Philosophies” (aidisruption.ai)
| Dimension | Value | Source |
|---|---|---|
| Release Date | 2026/8/14 0:00 UTC (8/14 12:00 GMT+8) | DeepSeek official + Black Whale WeChat |
| License | MIT (most permissive — commercial use + modification + closed-source derivatives) | DeepSeek official announcement |
| GitHub Stars (5 days cumulative) | ≈ 173,700 | Tencent News 8/22 |
| GitHub Forks (5 days cumulative) | ≈ 17,880 | Tencent News 8/22 |
| Hacker News Score | 744 points + 310 comments | Hacker News |
| Project Codename | DSH (DeepSeek Harness), black-whale WeChat branding (distinguished from the blue whale of model products) | DeepSeek official |
| Core Philosophy | ”Everything is a plugin” | DeepSeek official README |
| Current Status | Preview (developer preview) | DeepSeek official |
2.2 DeepSeek Harness vs. Codex Harness Engineering Philosophy Comparison
AI Disruption 8/24 “DeepSeek Harness vs. Codex Harness” deeply analyzes two Harness philosophies:
Core Assertion: “Agent = Model + Harness” has become industry consensus. Large models are engines; Harness is the engineering system that turns the engine into a controllable car that can actually run on the road. Codex Harness and DeepSeek Harness happen to represent two completely different “car manufacturing” philosophies.
| Dimension | DeepSeek Harness (DSH) | OpenAI Codex Harness |
|---|---|---|
| License | MIT (most permissive) | Apache-2.0 |
| Release Date | 8/14 | 8/19 |
| 5-Day Stars | 173,700 | 107,400 (as of 8/21) |
| Embeddability | Fully modular + self-hostable | Production-environment embed-friendly |
| Core Stack | MIT license, swap any model adapter | Rust core (codex-rs) + TypeScript SDK dual-stack |
| Monetization | Open Harness + raise V4 Pro API 2-4x (off-peak $1.98/M, peak $3.96/M, original $0.87/M) | Harness open-source + Codex/ChatGPT Work/Pro subscription |
| Strategy | ”Open the stack, raise the toll" | "Codex as a platform” |
Key Insight: On 8/16 DeepSeek simultaneously introduced peak/off-peak pricing (V4 Pro output from $0.87/M → off-peak $1.98/M + peak $3.96/M, 2.3-4.6x increase), while MIT-opening Harness — developers who find the API too expensive can self-host V4 Pro model + run DSH locally. This is the dual-sided “open the stack layer + raise the toll on the hosted layer” strategy.
III. Bullet 2: 8/19 OpenAI Codex Harness Apache-2.0 Fully Open-Sourced — 107,443 Stars + 3-Tier Integration Interfaces + “Codex as a Platform”
3.1 Key Hard Data
Source: OpenAI Official Blog “Codex as a platform: build on the open agent harness” (8/19) + Qiniu Cloud 8/21 “Codex Harness Full Open-Source: OpenAI Opens Agent Runtime Underlying Layer to Developers, Complete Analysis of Three-Tier Integration Interface” (news.qiniu.com/archives/1787276125198) + Weibo 8/21 “OpenAI Open-Sources Codex Harness: Agent Competition Center of Gravity Shifts from Model Layer to Execution Layer” (weibo.com/ttarticle/p/show?id=2309405334299804565787) + Tencent News 8/22
| Dimension | Value | Source |
|---|---|---|
| Release Date | 2026/8/19 OpenAI official blog | OpenAI official |
| License | Apache-2.0 | OpenAI official |
| Repository | github.com/openai/codex | OpenAI official |
| Stars (as of 8/21) | 107,443 | Qiniu Cloud 8/21 |
| Forks (as of 8/21) | 16,354 | Qiniu Cloud 8/21 |
| Latest Stable Version | v0.149.0 (released 8/20) | Qiniu Cloud 8/21 |
| First Complete Open-Source | Rust core codex-rs + app-server driver layer (previously only CLI frontend open-sourced) | Qiniu Cloud 8/21 |
3.2 OpenAI Codex Harness 3-Tier Integration Interfaces
Source: Qiniu Cloud 8/21 complete analysis + Weibo 8/21 + OpenAI official blog
OpenAI official positioning shift: “your application owns product context, business rules, and tools; Codex app-server provides the agent loop.” — Teams don’t have to move all work into a universal programming assistant; they can put Agents into software designed around real tasks (engineering workflows, operations panels, security investigations, customer support consoles).
| Interface Tier | Name | Use Case | Protocol Layer |
|---|---|---|---|
| Tier 1 (Light Non-Interactive) | codex exec | CI/CD pipelines, batch scripts, no human intervention needed | CLI |
| Tier 2 (Programmatic Orchestration) | Codex SDK | Custom Agent workflows, embedding enterprise SaaS | TypeScript SDK |
| Tier 3 (Persistent Session Driver) | Codex app-server | Long-running sessions, threads, turns, event streams, approval protocols | HTTP/JSON-RPC |
3.3 OpenAI Engineer Dominik Kundel’s Key Judgment
Source: AI Engineer World Expo, OpenAI engineer Dominik Kundel:
“Inference is no longer the bottleneck; the bottleneck is actually the network.” When model generation speed is fast enough, what truly constrains Agent capability is how it acquires context, calls tools, handles long tasks, and runs safely within boundaries. All these issues fall onto the Harness layer.
IV. Bullet 3: Microsoft Agent Lightning v1.0 — 6000 Samples Lifts Qwen3.5-9B SWE-bench +14.6 + Agent Lightning Skill
4.1 Key Hard Data
Source: cnblogs 8/25 “AI Tech Daily - 2026-08-25” (cnblogs.com/itech/p/22667102) + GitHub microsoft/agent-lightning/releases/tag/v1.0.1
| Dimension | Value | Source |
|---|---|---|
| GA Release Date | 2026/8/24 v1.0 + 8/25 v1.0.1 (Skill edition) | Microsoft Agent Lightning official |
| Project Stars | 17,600 | GitHub |
| Core Code Volume | ≈ 3,500 lines (after refactor simplification) | cnblogs 8/25 |
| Core Capability | ”Use production-time real Agent Harness for RL training” (Agent interacts with model via framework proxy, tools, context, control flow, and environment all stay in-loop, no code changes needed) | Microsoft official documentation |
| Official Demo Result | 6,000 training samples + Qwen3.5-9B programming workflow → SWE-bench Verified 41.8% → 56.4% (+14.6) | Microsoft Agent Lightning official |
| v1.0 Bundled Capability | Native Kubernetes Job-based Agent execution | Microsoft official |
| v1.0.1 New Capability | Agent Lightning Skill — lets Claude Code, Codex, GitHub Copilot etc. programming Agents systematically optimize other AI Agents’ prompts, tools, and workflows | Microsoft official |
| Open Source | Data cleaning, anti-reward-hacking, training scripts all open-sourced | Microsoft official |
4.2 Strategic Significance
Agent Training Paradigm Shift: Traditional RL training requires detaching from deployment environment and modifying code (“Harness is fixed”); Agent Lightning v1.0 allows directly running RL in Agent’s existing Harness — this means Agent can “self-evolve on the same Harness used in production”, and SWE-bench Verified +14.6 percentage points is just the beginning.
V. Bullet 4: Omarchy 4.0 “AI Operating System” — DHH + Dell + Dorsey + 5 Co-Founders 8-Person Omacom Foundation 8M USD
5.1 Key Hard Data
Source: Toutiao 8/25 “AI Internet Daily: Xiaomi On-Device AI Hardware Debut / ByteDance Integrates TRAE and Coze / Huawei Tri-Fold Launch Event Confirmed” (toutiao.com/article/7677806132395639306) + Tencent Securities 8/25 “Tencent Research Institute AI Express 20260825” (gu.qq.com/resources/shy/news/detail-v2/index.html)
| Dimension | Value | Source |
|---|---|---|
| Omarchy Version | 4.0 (Linux distro with AI coding Agent built directly into the system) | Tencent Securities 8/25 |
| Interaction Paradigm | One natural-language sentence to operate the computer (all system components become objects AI can read + modify) | Tencent Securities 8/25 |
| Community Contributions | 10 days 1,000+ plugin contributions | Tencent Securities 8/25 |
| Architecture Rewrite | System desktop rewritten as single process + all configuration changed to text scripts | Tencent Securities 8/25 |
| Foundation | Omacom Foundation | Tencent Securities 8/25 |
| Founders | DHH (David Heinemeier Hansson, Ruby on Rails author) + Dell (1 representative) + Jack Dorsey (Block/Twitter founder) + 5 co-founders, total 8 people | Tencent Securities 8/25 |
| First-Round Funding | 8M USD | Tencent Securities 8/25 |
| Strategic Positioning | ”New OS-layer opportunity above models, below hardware” | Tencent Securities 8/25 |
5.2 Omacom Foundation “Operating System = Harness” Core Assertion
“Harness used to serve the model company’s own Agent products; why are top model companies starting to actively open Harness?” Underlying these questions, AI’s main battlefield is now extending to “Agent runtime”. ——Tencent News 8/22 “After Black Whale, OpenAI Opens Harness”
Omarchy 4.0 takes this logic to the extreme: the operating system itself becomes a Harness — all configuration files are text, every system component can be read + modified by AI, which equals elevating Harness to the OS layer. Omacom Foundation’s 8M USD first-round funding bets on “a new layer above models, below hardware”.
VI. Bullet 5: xAI Grok Bot 8/22-23 Launches 24/7 Autonomous Cloud Agent — Always-on Agent Teammate Paradigm
6.1 Key Hard Data
Source: HeadsUpAI 8/23 “xAI Launches Grok Bot Beta for Autonomous AI Teammate Workflows” + HeadsUpAI 8/23 “xAI Expands Grok Bot Access to Cursor and SuperGrok Subscribers” (headsupai.io/ai-news-and-updates)
| Dimension | Value | Source |
|---|---|---|
| Release Date | 2026/8/22-23 Beta launch | xAI official |
| Core Paradigm | 24/7 autonomous cloud AI teammate — persistent AI teammate signs into tools to autonomously complete multi-step tasks on its own cloud computer 24/7 | xAI official |
| Subscriber Tier | SuperGrok Heavy + Cursor Ultra + Cursor Teams Premium (subscription-included) | xAI official |
| Cross-Platform | Desktop + iOS | xAI official |
| Learning Capability | Learns routines from user workflows + collaborates in shared threads | xAI official |
| Cursor Sync Access | SuperGrok Plus + Cursor Pro+ + Cursor Teams all automatically subscription-active | xAI official |
| Free Trial | All other users: limited free trial | xAI official |
6.2 Connection with the “Harness” Paradigm
Grok Bot’s core mechanism: “always-on AI agent service” + “sign into tools to complete multi-step tasks autonomously” + “on their own cloud computers” — this is the inflection point where the Harness layer upgrades from “development framework” to “permanent online SaaS”. OpenAI engineer Dominik Kundel’s “inference is no longer the bottleneck; the bottleneck is actually the network” is validated once again here.
VII. Bullet 6: 8/23-24 ByteDance TRAE + Coze Merge into Doubao — First Time a Chinese Internet Giant Fully Integrates Agent Product Lines
7.1 Key Hard Data
Source: Toutiao 8/25 “AI Internet Daily” + so.html5.qq.com 8/25 “ByteDance Integrates AI Product Lines: TRAE and Coze Merge into Doubao, Xiaomi XRZ Chip Breaks Through On-Device AI” (so.html5.qq.com/page/real/search_news?docid=70000021_5126a8d17ce58852)
| Dimension | Value | Source |
|---|---|---|
| Integration Timeline | 8/23 first reported → 8/24 +2 follow-ups → 8/25 Toutiao 12:19 summary | Toutiao 8/25 |
| Merged Products | TRAE Work (programming IDE) + Coze (office capability) | Toutiao 8/25 |
| Retained Product Line | TRAE IDE + CLI (continue as Doubao’s programming products) | Toutiao 8/25 |
| Unified Reporting | Zhao Qi (Doubao head) | Toutiao 8/25 |
| New Brand | ”Doubao Work” (ByteDance fastest this week launches independent AI office product + brand) | Toutiao 8/25 |
| Earlier Integration | End of July Feishu product team already merged with Doubao product team + Feishu aily renamed “Doubao Work Partner” | Toutiao 8/25 |
7.2 Strategic Significance
Three-Combo Hit: (1) End of July Feishu + Doubao product teams merged; (2) Early August Feishu aily renamed “Doubao Work Partner”; (3) 8/23-24 TRAE Work + Coze fully merged into Doubao. This is the first time a Chinese internet giant fully integrates Agent product lines — meaning “single-vendor Agent platform” is moving toward “cross-product-line Agent collaboration”, making ByteDance Doubao the “China Microsoft Copilot + ChatGPT Work + Agent Builder” three-in-one.
VIII. Same Period: 8/23 Hugging Face Explores 13B USD Sale — AI Middle Layer Being Bought Up + Open-Source Community Neutrality Concerns
8.1 Key Hard Data
Source: Tencent News 8/24 “GitHub AI Daily” (new.qq.com/rain/a/20260825A021QO00?refer=cp_1009) + Tencent Securities 8/25 + Toutiao 8/25
| Dimension | Value | Source |
|---|---|---|
| Disclosure Date | 2026/8/23 foreign media | Tencent News 8/24 |
| Valuation | 13B USD | Tencent News 8/24 |
| 2023 Valuation Comparison | 4.5B USD → +189% | Tencent News 8/24 |
| Company | Hugging Face (founded 2016, world’s largest AI community) | Tencent News 8/24 |
| Platform Scale | 3M+ models + 1M+ datasets | Tencent News 8/24 |
| Acceleration Reason | July security incident → accelerated monetization | Tencent News 8/24 |
| Comparison | Stripe in August just acquired model routing platform OpenRouter at over 8B USD | Tencent News 8/24 |
8.2 Three-Layer Signal
Signal 1 (Middle Layer Valuation): “AI value is rapidly spreading to middle-layer platforms beyond models” — Stripe acquired OpenRouter (routing layer) + Hugging Face explored sale (community/distribution layer). Hugging Face doesn’t build models, but becomes the developer distribution entry and the “water, electricity, coal” of open-source AI ecosystem.
Signal 2 (Open-Source Community Concerns): Once Hugging Face is acquired by a cloud vendor, open-source community worries whether neutrality can be maintained — “super-large models” and “super-large community” belong to different entities, and open-source engine + closed-source channel is very likely to emerge.
Signal 3 (Capital Flow): Agent Harness open-source 8/14-19 + middle-layer platform acquisition 8/23-24 — AI middle layer (middleware, community, routing, Harness) is being densely bought up by capital.
IX. Judgment: “Agent = Model + Harness” Industry Consensus + 5-Step Enterprise Landing + 6-Layer Defense Checklist
9.1 Consensus Upgrade Timeline
| Time | Event | ”Agent = X” Consensus Upgrade |
|---|---|---|
| 2024 | AutoGPT / BabyAGI single Agent | Agent = LLM + Prompt + Tools |
| 2025 H2 | LangGraph / CrewAI | Agent = LLM + Workflow Engine |
| 2026 Q2 | MCP / OpenAI Agent Plugins | Agent = LLM + Standardization Protocol |
| 2026 Q3 | Codex Harness + DSH + Agent Lightning + Omarchy | Agent = LLM + Harness (execution layer + training layer + OS layer three-axis resonance) |
| 2026 August | Grok Bot + ByteDance Doubao merger + Hugging Face 13B | Agent = LLM + Harness + 24/7 Service + Product Line Integration + Middle-Layer Capital |
9.2 5-Step Enterprise Landing Path
| Step | Action | Time Window | Dependency |
|---|---|---|---|
| Step 1 | Inventory existing Agent Harness — assess whether existing CLI, SDK, automation scripts constitute a Harness | 1 week | Engineering inventory |
| Step 2 | Pilot DeepSeek Harness (DSH) or Codex Harness 3-tier interfaces | 2 weeks | Sandbox + CI/CD |
| Step 3 | Use Agent Lightning v1.0 to run RL on production Harness (+14.6 SWE-bench verified) | 4 weeks | Compute resources + RLHF team |
| Step 4 | Integrate Always-on Agent Teammate (reference xAI Grok Bot 24/7) | 6 weeks | Persistent session + approval protocol |
| Step 5 | Platform-level integration (reference ByteDance Doubao model: unified reporting + cross-product-line Agent) | 8 weeks | Product lines + brand + reporting-line integration |
9.3 6-Layer Defense Checklist
| # | Defense Item | Key Approach |
|---|---|---|
| 1 | Harness sandboxing | All Agent tool calls go through sandbox (Codex app-server sandbox config can be referenced) |
| 2 | Approval protocol logging | Critical operations go through human approval flow (e.g. Codex app-server approval event flow) |
| 3 | Training data cleaning + anti-reward-hacking | Reference Agent Lightning v1.0 anti-reward-hacking open-source scripts |
| 4 | Harness multi-stack compatibility | Support at least MIT and Apache-2.0 two open-source Harnesses (DSH + Codex) |
| 5 | 24/7 Agent monitoring | Always-on Agents require continuous monitoring (reference xAI Grok Bot shared thread collaboration) |
| 6 | Platform neutrality contingency | Contingency plan for open-source model distribution channels once Hugging Face is acquired |
FAQ (High-Frequency Questions Direct Answer)
Q1: What is Harness? Why is it more important than the model? A: Harness is the underlying execution system that drives Agent operation — managing conversation state, tool calls, sandbox execution, approval mechanisms, context compression, multi-turn task progression. A capable Agent cannot have just Prompt and model response; it must run a controllable, approvable, persistable execution system around the model. OpenAI engineer Dominik Kundel: “Inference is no longer the bottleneck; the bottleneck is actually the network.” All these issues (calling tools / handling long tasks / running safely within boundaries) fall onto the Harness layer.
Q2: What is the essential difference between OpenAI Codex Harness and DeepSeek Harness (DSH)? A: DSH (MIT, most permissive) leans toward “fully modular + self-hostable + Everything is a plugin”; Codex Harness (Apache-2.0) leans toward “production-environment embed-friendly + Rust + TypeScript SDK dual-stack”. DSH leaves Monetization to hosted API (V4 Pro output price 2-4x increase); Codex Harness leaves Monetization to Codex/ChatGPT Work subscription. Underlying philosophy difference: DSH = “Open the stack, raise the toll”; Codex = “Codex as a platform”.
Q3: What is the biggest breakthrough in Microsoft Agent Lightning v1.0? A: 6,000 training samples lift Qwen3.5-9B SWE-bench Verified from 41.8% to 56.4% (+14.6), no need to modify any existing Agent code — Agent interacts with model via framework proxy, tools, context, control flow, and environment all stay in-loop. Greatest significance: Agent can “self-evolve on the same Harness used in production”, and SWE-bench +14.6 is just the beginning.
Q4: What is the difference between Omarchy 4.0 and traditional Linux distros? A: Traditional distros have GUI processes + binary configuration; Omarchy 4.0 rewrites the system desktop as a single process + changes all configuration to text scripts, every component can be read + modified by AI — “the operating system itself becomes a Harness”. DHH + Dell + Jack Dorsey etc. 8 people formed Omacom Foundation + 8M USD first-round funding, betting on “new OS-layer opportunity above models, below hardware”.
Q5: What is the difference between xAI Grok Bot and traditional Chatbots? A: Grok Bot is a 24/7 autonomous cloud AI teammate that autonomously manages tool flows 24/7 on independent cloud computers, learns routines from user workflows, and collaborates in shared threads; Chatbots passively respond to user queries. Grok Bot is integrated into SuperGrok Heavy + Cursor Ultra + Cursor Teams Premium subscriptions — this is the inflection point where the Harness layer upgrades from “development framework” to “permanent online SaaS”.
Q6: What does ByteDance’s TRAE + Coze merger into Doubao mean for the Chinese AI industry? A: It means first time a Chinese internet giant fully integrates Agent product lines — end of July Feishu + Doubao product teams merged + Feishu aily renamed “Doubao Work Partner” + 8/23-24 TRAE Work + Coze fully merged into Doubao + ByteDance fastest this week launches “Doubao Work” independent brand. Single-vendor Agent platform moving toward cross-product-line Agent collaboration, Doubao becomes China’s Microsoft Copilot + ChatGPT Work + Agent Builder three-in-one.
Q7: What does Hugging Face’s exploration of 13B USD sale mean? A: It means AI middle layer is being densely bought up by capital — Stripe in August acquired model routing platform OpenRouter at over 8B USD, Hugging Face explores 13B sale (2023 valuation 4.5B → +189%). AI value is rapidly spreading to middle-layer platforms beyond models (routing, community, Harness). At the same time open-source community worries about neutrality — once Hugging Face is acquired by a cloud vendor, “open-source models + closed-source channels” pattern may take shape.
Q8: Which Harness route should enterprises choose at the current stage? A: Three-axis parallel: (1) Deep integration layer — choose Codex Harness 3-tier interfaces (production-environment embed-friendly); (2) Autonomous controllable layer — choose DeepSeek Harness (DSH) MIT self-hosting; (3) Agent training layer — use Microsoft Agent Lightning v1.0 to run RL on production Harness. Simultaneously prepare for Omarchy 4.0-class AI OS + Grok Bot-class Always-on Agent pilots.
Q9: What is the difference between “Agent = Model + Harness” and the past “Agent = LLM + Workflow Engine”? A: The past “workflow engine” is fixed orchestration (nodes + edges + DAG); Harness is the sum of “all boundaries + state + context of Agent-world interaction”, including sandbox, approval, long-task orchestration, context compression, multi-Agent collaboration, persistent sessions. Core difference: Harness is “the deciding factor of how far Agent can run in production”; the model is just an engine.
Q10: How will the Harness track evolve in the next 6 months? A: Three-axis prediction: (1) Model-layer giants continue opening Harness (OpenAI → Anthropic Claude Code partially opened in August → Google Gemini CLI follows); (2) Middle-layer platforms continue being acquired (Hugging Face 13B → Cohere / Together AI follow); (3) AI OS becomes a new track (Omarchy 4.0 → Apple Intelligence 2.0 + Windows AI Foundry follow).
Key Terminology
| Term | Explanation |
|---|---|
| Harness | Underlying execution system that drives Agent operation — sum total of managing conversation state, tool calls, sandbox execution, approval mechanisms, context compression, multi-turn task progression |
| Agent = Model + Harness | 2026 Q3 industry consensus — large models are engines, Harness is the engineering system that turns the engine into a controllable car that can actually run on the road |
| Apache-2.0 | Open source license adopted by OpenAI Codex Harness — allows commercial use + modification + closed-source derivatives, slightly stricter than MIT (retains patent license) |
| MIT License | Open source license adopted by DeepSeek Harness (DSH) — most permissive, allows commercial use + modification + closed-source derivatives, no restrictions |
| codex exec | OpenAI Codex Harness Tier 1 interface — light non-interactive calling, suitable for CI/CD pipelines |
| Codex SDK | OpenAI Codex Harness Tier 2 interface — TypeScript SDK, supports programmatic orchestration of custom Agent workflows |
| Codex app-server | OpenAI Codex Harness Tier 3 interface — HTTP/JSON-RPC persistent session driver, supports threads, turns, event streams, approval protocols |
| Agent Lightning v1.0 | Microsoft 8/24-25 released Agent RL training framework — no need to modify any existing Agent code, run RL on production Harness |
| Omarchy 4.0 | Linux distro initiated by DHH etc. — directly embeds AI coding Agent into operating system, “above models, below hardware” |
| Grok Bot | xAI 8/22-23 launched 24/7 autonomous cloud Agent — autonomously manages tool flows 24/7 on independent cloud computers, learns routines from user workflows |
| DSH (DeepSeek Harness) | DeepSeek 8/14 MIT open-source Agent Harness — “Everything is a plugin” philosophy, black-whale WeChat branding |
| Omacom Foundation | AI Operating System foundation established by DHH + Dell + Jack Dorsey etc. 8 people — 8M USD first-round funding |
| Agent Runtime | Key infrastructure layer for how far Agent can run in production environment — contains sandbox, approval, context, long-task orchestration, persistent sessions |
| Always-on Agent Teammate | 24/7 persistent-online Agent service model — pioneered by Grok Bot, redefines Harness from SaaS perspective |
References
I. Official Announcements and Original Sources (6 entries)
- OpenAI Official Blog: “Codex as a platform: build on the open agent harness” (2026/8/19), github.com/openai/codex Apache-2.0 license open-sourced, 107,443 Stars + 16,354 Forks, v0.149.0
- DeepSeek Official Announcement: “DeepSeek Harness (DSH) MIT Open-Source” (2026/8/14), black-whale WeChat branding, 5 days 173,700 Stars + 17,880 Forks, HN 744 points 310 comments
- Microsoft GitHub Release: “microsoft/agent-lightning/releases/tag/v1.0.1” (2026/8/24-25), v1.0 GA + v1.0.1 Skill, 17,600 Stars
- Toutiao 8/25 “AI Internet Daily”: toutiao.com/article/7677806132395639306, Omarchy 4.0 + ByteDance TRAE merger into Doubao + Xiaomi XRZ + Alibaba Wan3.0 + NVIDIA Vera Rubin
- Tencent Securities 8/25: gu.qq.com/resources/shy/news/detail-v2/index.html?t=1#/index?_tentrees_trans=0&id=SN20260825000905b6a01849, “Tencent Research Institute AI Express 20260825”
- xAI Official Announcement (8/22-23): Grok Bot Beta for Autonomous AI Teammate Workflows, integrated into SuperGrok Heavy + Cursor Ultra + Cursor Teams Premium
II. In-Depth Technical Analysis (5 entries)
- Qiniu Cloud 8/21: “Codex Harness Full Open-Source: OpenAI Opens Agent Runtime Underlying Layer to Developers, Complete Analysis of Three-Tier Integration Interface”, news.qiniu.com/archives/1787276125198
- Weibo 8/21: “OpenAI Open-Sources Codex Harness: Agent Competition Center of Gravity Shifts from Model Layer to Execution Layer”, weibo.com/ttarticle/p/show?id=2309405334299804565787
- Tencent News 8/22: “After Black Whale, OpenAI Opens Harness, Competing for ‘Agent Runtime’”, news.qq.com/rain/a/20260822A027E700?adChannelId=tech
- AI Disruption 8/24: “DeepSeek Harness vs. Codex Harness: A Clash of Engineering Philosophies”, aidisruption.ai/p/deepseek-harness-vs-codex-harness
- Studio Global AI 8/24: “DeepSeek Harness v0.1: The Open-Source Agent Framework That Broke GitHub Records”, studioglobal.ai
III. Industry Daily and Media (6 entries)
- cnblogs 8/25: “AI Tech Daily - 2026-08-25” Top 10, cnblogs.com/itech/p/22667102 (Microsoft Agent Lightning v1.0 + LLM inference engine vulnerabilities)
- Tencent News 8/24: “GitHub AI Daily”, new.qq.com/rain/a/20260825A021QO00?refer=cp_1009 (Hugging Face explores 13B USD sale + Alibaba Wan3.0 + OpenAI gpt-nathree leak)
- so.html5.qq.com 8/25: “ByteDance Integrates AI Product Lines: TRAE and Coze Merge into Doubao, Xiaomi XRZ Chip Breaks Through On-Device AI”, so.html5.qq.com/page/real/search_news?docid=70000021_5126a8d17ce58852
- Weibo 8/24-25: “2026/8/25 AIGC Industry News Express” (NVIDIA Nemotron 3.5 Lightning + xAI Grok 4.6 + Thomson Reuters foundation model), weibo.com/7905315703/5335726201769542
- Weibo 8/25: “AI Developer Daily · 0825”, Doubao TRAE + Coze + Xiaomi XRZ + Alibaba Wan3.0 + Multi-Model Agent
- HeadsUpAI 8/25: “Trending AI News Past Week (Aug 18-25, 2026)” Top 30 of 144 (Grok Bot x2 + Nous Research Ox Alpha + Generalist GEN-1.5 + OpenAI GPT-5.6 Sol 20% price cut + Unsloth Qwen3.8 + Anthropic Claude Code 2.1.237), headsupai.io/ai-news-and-updates
IV. Models and Ecosystem (3 entries)
- Toutiao 8/25: “AI Internet Daily” on TRAE integration + Doubao Work independent brand + Xiaomi XRZ O100
- Tencent Securities 8/25: “Tencent Research Institute AI Express” on Anthropic marshmallow/melon models + Omarchy 4.0 Foundation + Hugging Face 13B
- DeepSeek Official V4 Pro API Pricing: 8/16 peak/off-peak pricing takes effect — output off-peak $1.98/M, peak $3.96/M (original $0.87/M), 2.3-4.6x increase
Written 2026-08-25 13:00 GMT+8 · Rujuan · 13:00 cron fallback manual backup (6:00 daily-publish cron failed 28th consecutive time, MiniMax FailoverError: AI service temporarily overloaded, consecutiveErrors accumulated 28, see topics.md change log)