Harness layer becomes $1.5T battleground as AI shifts from models to delivery

The race to dominate AI is no longer about who builds the smartest model — it is about who can reliably deliver the result.

Alibaba, Tencent, OpenAI and xAI are racing to integrate the Harness layer — the middleware that sits between AI models and end users — as competition shifts from raw model capability to task delivery, a shift that expands the addressable market from roughly $200 billion to $1.5 trillion, according to a CITIC Securities report published this week.

"The model determines the ceiling of intelligence, but the Harness determines whether that intelligence can be stably, controllably and cost-effectively transformed into real results," the CITIC Securities analysts wrote, arguing that the layer's value lies in lowering usage barriers and capturing high-quality long-range task data.

The urgency stems from a fundamental gap between single-step capability and end-to-end delivery. Claude Mythos Preview's success rate drops to 50% on tasks lasting about 17 hours, according to METR, while Anthropic estimates multi-agent systems consume roughly 15 times the tokens of standard chat sessions. Microsoft's CodeAct Harness, by contrast, cut execution time by 52.4% and token consumption by 63.9% through optimized tool orchestration, demonstrating that the execution layer — not the model — often determines whether complex tasks complete.

The Harness layer connects upstream model supply to downstream customer demand, routing requests across multiple models, managing context, memory, tool calls, permissions and state, and delivering the final output. By replacing the specialized integrated development environment with a natural-language office interface, the layer expands the addressable market from roughly $200 billion in AI coding to $1.5 trillion in general white-collar knowledge work, the CITIC report estimates.

The Four-Phase Pipeline That Makes Models Deliver

Alibaba consolidated QoderWork, Wukong and MuleRun into a single product called Qwen Work. Tencent pushed its WorkBuddy and QClaw teams to converge. OpenAI deepened the integration between ChatGPT and Codex. xAI strengthened its connection to Cursor. Each move targets the same structural bottleneck: models can reason, but they cannot reliably execute.

The Harness layer operates across four phases — code ingestion, threat modeling, deep-dive verification and exploit chain synthesis — before reaching remediation and fix validation, according to Visa's Project Glasswing white paper, which the payments giant published after using Anthropic's Claude Mythos to hunt vulnerabilities in its own network. Visa open-sourced its Vulnerability Agentic Harness on GitHub, where it has collected 595 stars and 97 forks as of July 20. The company also introduced a new metric — Mean Time to Adapt — that measures how quickly a team can confirm an exploit, fix it and prove the attack path is closed, replacing traditional mean-time-to-detect metrics that Visa argues can mask growing exposure.

Data Reflux Creates a Self-Reinforcing Moat

The strategic value of the Harness extends beyond execution reliability. Each completed task generates a trajectory — the sequence of tool calls, error corrections and decision points — that becomes training data for the next model iteration. This creates what CITIC calls a "flywheel" of TAM expansion, data reflux and product enhancement that becomes harder for latecomers to replicate.

China's desktop office agent market illustrates the acceleration. Monthly visits grew from more than 20 million in March 2026 to more than 60 million in June, according to Analysys, as the bottleneck shifted from "can it be used" to "can it be delivered stably and at scale." WorkBuddy and similar products now offer multi-model routing, weakening the link between any single model provider and the end user, while capturing the user relationship, task history and feedback data in the Harness layer.

The implication for investors is that the moat in AI is migrating from model weights to the execution layer. As open-source models close the capability gap — Cisco's Antares-1B reached 0.209 File F1 on vulnerability localization, beating GLM-5.2 at 753 billion parameters — the ability to route tasks to the right model, manage state across long-running agents and capture the resulting data becomes the durable competitive advantage. Companies with mature Harness products, cloud infrastructure and high-frequency office entry points — Alibaba, Tencent, and platform providers with similar assets — are best positioned to capture the $1.5 trillion addressable market, according to CITIC. The report recommends prioritizing firms that have already demonstrated a closed loop from unified entry point to task execution to result delivery.

This article is for informational purposes only and does not constitute investment advice.