To The Moon

TMTPOST — For the past year, the global artificial intelligence boom has lived largely in the consumer cloud—a world of conversational novelties, fleeting viral interactions, and generative text experiments. But inside the high-tech corridors of China’s leading enterprise software providers, the romantic era of the broad-spectrum large language model (LLM) is giving way to a much more calculating, corporate reality.

On July 10, 2026, Baidu formally advanced its general-purpose intelligent agent product, "Baidu Dazi," into the enterprise market. The transition moves the platform beyond its origins as an individual office assistant and into a structured corporate agent framework designed to anchor internal workflows.

The expansion comes as the technology sector experiences an industry-wide pivot. As basic model parameters and raw reasoning capabilities face plateauing returns or unsustainable capital expenditures, competition has shifted toward AI Agents—autonomous entities capable of executing multi-step tasks across complex digital environments. The enterprise, with its deep pockets and predictable inefficiencies, has become the primary battleground. According to Shen Dou, Executive Vice President of Baidu and President of the Baidu Smart Cloud Group, the trajectory is clear: "In the future, 90% of work could involve deep participation from intelligent agents."

Yet, moving an AI from a casual chat window into the highly regulated, brittle architecture of a modern corporation is revealing a deep friction between theoretical model capabilities and practical engineering execution.

The Engineering Realities of Autonomy

When an individual user asks an AI to draft a generic email, an unexpected error or hallucination is an annoyance. When an enterprise agent is tasked with cross-referencing an inventory database, generating a purchase order via an ERP system, and executing a payment workflow, an unexpected error is a liabilities nightmare.

To bridge this gap, Baidu's enterprise push includes the integration of foundational business infrastructure: data asset accumulation, multi-user collaboration tools, and direct linkages to legacy internal systems like OA (Office Automation), CRM (Customer Relationship Management), and ERP. Crucially for enterprise compliance, it introduces employee identity pass-through and row-level data permission management—protocols designed to ensure an AI cannot inadvertently expose sensitive financial or HR data to unauthorized staff.

Furthermore, Baidu introduced the industry's first enterprise-grade Skill access standard, aiming to establish a unified system for third-party integrations.

The necessity of these rigid structures points to a broader technical challenge. "Compared to the era of simple chatbots, intelligent agents must handle more complex reasoning, multi-tool orchestration, massive file reading and writing, as well as permissions, security, cost, and data governance issues," Shen Dou observed. Managing this requires an integrated, full-stack approach that spans computing infrastructure, model scheduling, and dedicated agent frameworks.

The 56% Problem: Where AI Actually Fails

For months, the market assumed that the primary bottleneck to effective AI deployment was the underlying model itself—the need for more parameters, more data, and higher compute. However, internal telemetry from deployment cycles suggests a different structural reality.

Li Jingqiu, Chief Product Architect of Baidu Dazi, revealed a revealing metric from internal testing: among instances where intelligent agents failed or delivered poor results, roughly 44% of the failures could be attributed directly to the limitations of the underlying LLM. The remaining 56% of failures stemmed from the capabilities of the "Harness" engineering framework—specifically its capacity for workload scheduling and intent understanding.

This data suggests that at the current stage of enterprise adoption, an agent platform's success relies less on the raw intelligence of the model and more on the precision of its surrounding plumbing. It requires robust frameworks capable of intent clarification, precise tool invocation, multi-round searching, and—critically—automated error correction when a third-party application throws an exception.

"Of course, large model capability remains highly important," Li noted. "The critical piece is the collaborative optimization between the LLM and the Harness framework."

Standardizing the "Skill" Economy

As tech giants compete to monetize through token consumption, the "Skill" ecosystem—the functional plugins that allow an AI to interact with the outside world—has become a messy frontier. Over the past few months, the uncontrolled growth of these ecosystems has exposed significant vulnerabilities: unverified security parameters, exaggerated capability descriptions, unstable invocation rates, and incompatible interface standards.

When third-party extensions are introduced into corporate networks, they represent an expanded attack surface. Baidu's response has been to prioritize verification over sheer ecosystem volume. "As the Skill ecosystem enters a phase of granular governance, ensuring the safety and reliability of invocations is far more important than scale," Li Jingqiu told reporters.

To mitigate these risks, the platform has deployed an automated security audit and validation mechanism. When a corporate user attempts to import an external, custom Skill, the system isolates and tests it within a secure software sandbox, automatically blocking detected anomalies.

Architecturally, the industry is coalescing around a three-tiered structural philosophy for these capabilities:

First, the Skill layer resolves long-tail, single-point functional demands. Second, the Connector layer enables bulk, application-level integrations, such as managing secure authentication protocols with enterprise communication tools like DingTalk or Lark. Finally, the Toolkit layer deploys specialized, professional suites that allow a worker to close the loop on an entire vertical workflow from a single input prompt.

"The professional toolkit phase represents a likely end-game for the ecosystem's development," Li remarked. "The goal is for every individual in the workplace to find their corresponding professional skill bundle."

The Path to Scale: Verticalization and ROI

Despite the rapid product rollouts from major tech firms, two distinct hurdles remain before intelligent agents achieve ubiquitous scale within the enterprise.

The first is the necessity of deep verticalization. General-purpose models understand language, but they do not intuitively understand the highly specific regulatory constraints of pharmaceutical compliance, the risk parameters of commercial underwriting, or the logistical nuances of supply-chain management. Overcoming this requires both specialized performance iterations at the model layer and the maturation of industry-specific toolkits.

The second, and perhaps most formidable, obstacle is hard economic calculation. The initial wave of corporate AI adoption was frequently funded by speculative innovation budgets. As those experimental cycles close, chief financial officers are demanding clear Return on Investment (ROI) metrics.

For enterprise agents to achieve friction-free adoption, the operational cost per token and the compute overhead of multi-step reasoning chains must drop significantly. Until the cost of deploying an autonomous agent is demonstrably lower than the human inefficiencies it seeks to correct, the corporate world's shift to AI agents will remain a disciplined, deliberate march rather than an overnight revolution.