By SOLOMOAT Editorial Team

Tencent Research Institute

Author: Liu Qiong | Originally published in Enterprise Management Magazine

Around the Lunar New Year of 2026, our day-to-day workflow underwent a decisive shift. Artificial intelligence transitioned from producing half-baked drafts requiring line-by-line human intervention into autonomous executors capable of delivering end-to-end assignments. The labor split between humans and machines inverted overnight from 90:10 to 20:8. This was not an incremental step; it arrived almost in a single leap.

💡 Core Strategic Takeaway

• Agentic AI transfers execution rights from employees operating software to systems acting on objectives.
• Management must shift from supervising tasks to defining goals, permissions, controls, and accountability.
• Enterprises that keep legacy structures will capture efficiency gains but miss the deeper organizational transformation.

Simultaneously, a broader transformation unfolded across the market. OpenClaw—an open-source, local device agent framework recognizable by its lobster logo—debuted on GitHub in early 2026. Within two months, its star count eclipsed that of the three-decade-old Linux kernel. Apple’s Mac Mini sold out globally as users rushed to acquire local hardware for OpenClaw. In the Nanshan technology hub of Shenzhen, hundreds of professionals—most outside of traditional software engineering—lined up to deploy similar agent architectures.

The shift from conversational interfaces to real execution gave everyday operators a visceral encounter with operational AI. Market consumption reflected this shift instantly: in early 2024, daily token consumption across China stood at roughly 100 billion; by March 2026, that figure hit 140 trillion, a 1,400-fold expansion in two years.

This surge signals more than an efficiency gain. It reflects a structural reallocation of operational authority: execution rights are transferring from human operators to autonomous agents.

The legacy internet paradigm operated on a direct chain: Human → Launch Software → Execute Task. The emerging paradigm operates on an abstracted model: Human → Define Objective → Agent Executes. Operational execution is shifting from humans managing software to software executing on behalf of humans (as seen in desktop agent solutions such as Tencent Cloud's WorkBuddy).

The Evolution of Agent Architectures

1. From Isolated Actions to Sustained Execution

Agent architectures have evolved through three distinct phases:

Phase One (Direct Tool Execution): Users provided explicit commands, and the system carried out isolated actions—translating a single passage or generating a flow chart. The human retained granular control over every step; the model merely operated faster. Early autonomous iterations in 2023, such as AutoGPT, attempted to leap directly to full autonomy, but frequently drifted off-course during extended execution chains, underscoring that capability upgrades require progressive validation.

Phase Two (Sandboxed Goal Planning): Users specified a project goal, and the agent planned and completed the intermediate milestones independently. For a competitive teardown, the system retrieved sources, synthesized data, ran cross-comparisons, formed conclusions, and structured the final deck. Offerings like Manus attained this standard. Because these models operated within cloud-based sandboxes without direct access to local enterprise files and native applications, their operational utility stopped short of real workspace environments.

Phase Three (Local, State-Persistent Execution): The paradigm shift of 2026 centers on local runtime execution. Platforms like OpenClaw run on local hardware, reading files, controlling terminals, operating web browsers, invoking desktop applications, and collaborating directly with operators via enterprise messaging channels such as WeCom. Crucially, these systems maintain persistent cross-task memory. They record operational habits, adapt to user preferences, accumulate execution history, and convert repetitive workflows into reusable internal assets.

Major Chinese technology providers moved quickly. Beginning in February 2026, Tencent, ByteDance, Alibaba, Baidu, and Xiaomi rolled out agent architectures and tailored enterprise solutions. On March 9, Tencent released two agent systems, WorkBuddy and QClaw; WorkBuddy saw first-day traffic shatter internal forecasts, prompting an immediate tenfold server expansion. ByteDance integrated its ArkClaw agent into the Feishu enterprise suite and upgraded Coze 2.0 into a long-horizon planning and cloud-execution engine.

Market strategies have split across three distinct entry points:

All three approaches target the same prize: capturing the user's operational trust. On the corporate side, research from iResearch indicates that over 78% of mid-to-large enterprises have integrated AI agents into core workflows. Moving from conceptualization to sustained local execution, agents are transitioning from basic productivity utilities into core execution engines.

2. The Determinant of Industrial Usability: The Harness Ecosystem

Model capability improvements are only the baseline. Deploying agents at scale depends less on foundational base models and more on the supporting systems engineering.

In 2024, only a handful of products could execute complex multi-step workflows, and model steering rarely extended beyond manual prompt design. By April 2026, leading agent architectures like OpenClaw, Hermes Agent, and Claude Code all converged on the classical loop (Plan → Execute → Verify → Iterate), yet diverged significantly in harness engineering. Some favor rigorous hierarchical controls; others emphasize autonomous skill compounding; others treat the underlying file system as infinite context memory.

As foundational model capabilities converge, competitive differentiation shifts to the harness layer—the systems engineering that coordinates tool invocation, context window management, dynamic memory retrieval, and workflow routing without altering base model parameters.

Consider an automotive parallel: the foundational model is the engine; when engine outputs equalize, actual performance is determined by transmission logic and chassis tuning. Technical benchmarks confirm that the same underlying model can exhibit double-digit percentage differences in task completion rates depending on the surrounding harness. As Dowson Tang, Senior Executive Vice President of Tencent, noted: "Enterprise AI adoption is not merely an algorithmic challenge; it is a systems engineering challenge."

Tencent unbundled these harness capabilities into modular infrastructure components—memory engines, storage tiers, and secure sandbox environments—enabling enterprises to deploy tailored systems without building custom execution stacks from scratch. When Tencent's open-source Agent Memory module was integrated with an off-the-shelf base model, long-horizon task completion rates rose from 33% to 50%, while token consumption dropped by 61%.

The developer ecosystem around these frameworks has expanded at a rapid pace. Agent Skills—the functional modular extensions for agents—emerged in late 2025 and surpassed 85,000 distinct modules by March 2026. Problems that previously required bespoke vertical software development can now be addressed by plugging in standardized skill modules. Tencent has packaged core functional capabilities across documents, enterprise conferencing, and digital mapping into over 35,000 standardized interfaces for dynamic agent orchestration.

Macroeconomic and Industry Implications

1. Machine Cognition as an Independent Factor of Production

Over the past three years, model inference costs fell by more than 90% annually. At the same time, open-source communities reduced deployment barriers from specialized engineering teams down to everyday operators. By early 2026, machine cognition shifted from an elite enterprise capability into an accessible, variable factor of production.

As an economic input, this execution capacity displays two structural traits: Accessibility and Measurability.

Accessibility: Costs and implementation thresholds have declined into ranges viable for mainstream enterprise budgets.

Measurability: Every invocation maps directly to discrete token consumption, allowing costs and outputs to be tracked, modeled, and optimized with precision.

Because execution output can be metered and billed by the token, enterprises can manage automated labor capacity much like physical raw materials. This marks a departure from traditional payroll economics. Human labor is tied to individual workers, making it difficult to unbundle or measure with surgical precision. Agent services scale dynamically with business volume and can be provisioned on demand.

As a result, an enterprise's operational bandwidth is no longer strictly bound to headcount, but to how effectively it orchestrates execution resources. Operational capacity has decoupled from individual employees into an independently allocatable asset, enabling corporate execution to be externalized, measured, and reallocated across business units.

Token volume alone does not confer a competitive edge. The decisive metric is conversion efficiency—the business yield generated per token consumed. Offloading execution to agents creates tangible economic value only when model activity converts reliably into strategic outcomes.

2. Structural Shifts in Cost Models and Software Procurement

Under legacy operational frameworks, scaling enterprise revenue required a parallel expansion in payroll. Execution relied on human labor, locking companies into rigid, fixed overhead: salaries, social insurance, and training expenses.

As agents absorb operational execution, these costs do not vanish; they change form: fixed payroll transforms into elastic compute and inference charges. Costs scale in direct alignment with market demand.

A 100-person market research firm previously needed 200 analysts to double its case throughput. By integrating agent workflows, that same 100-person core can deliver multiples of historical output.

Software procurement is experiencing an equivalent shift. Enterprises previously purchased static software licenses and seat-based access; today, pricing structures increasingly anchor to verified task outcomes.

Enterprise procurement is moving away from hiring heads or licensing interfaces toward directly contracting execution capacity. Market advantage will accrue to organizations that secure reliable execution with lower unit economics and greater operational stability.

Building Proprietary Enterprise Agents

To capture real strategic margin, organizations must move beyond generic agent platforms and construct proprietary enterprise agents. This requires fusing base models with internal operational playbooks, proprietary data assets, and years of firm-specific operating history.

The objective is simple: capture the implicit knowledge previously locked in the minds of veteran personnel, translate it into accessible system rules, and convert individual expertise into reusable corporate infrastructure.

Implementation paths vary by corporate scale. Large enterprises will deploy fully custom stacks, whereas small and mid-sized operators can move faster via specialized SaaS integrations and modular skill packs. The foundational rollout sequence remains identical: begin with a high-frequency, standardized, fault-tolerant workflow before expanding horizontally across the organization.

1. Converting Institutional Knowledge into Systematic Assets

Proprietary agents differ from generic foundation models because they are trained on internal operating realities. They institutionalize the unwritten insights scattered across senior staff, internal documentation, and historical transaction records into actionable knowledge assets.

Target High-Frequency Workflows: In furniture export manufacturing, determining margin concessions on complex custom quotes historically required direct apprenticeship under veteran sales directors, taking junior staff six months to reach basic competence. By translating pricing boundary conditions into deterministic logic matrices for an agent, entry-level staff can execute quotes on day one with accuracy on par with multi-year veterans. Similarly, property management firms that embed tenant ticket routing and priority rules into agents can automatically categorize 80% of customer issues and generate resolution paths, leaving human managers to focus on the 20% of edge cases requiring subjective judgment.

Translate SOPs into Machine-Executable Logic: Most enterprises maintain standard operating procedures, but these documents are written for human interpretation. Agents require precise execution criteria: exact conditions that trigger actions, specific parameters that demand human escalation, and clear boundaries for full autonomy. This translation process systematically surfaces unaddressed operational ambiguities that previously relied on informal staff assumptions to function.

Build Continuous Feedback Loops: A proprietary agent is not a static installation. Every successful run reinforces optimal execution paths, while every human correction refines operational boundaries. After six months of live operation, an agent's grasp of unique enterprise context becomes an asset that competitors cannot replicate simply by purchasing off-the-shelf software. The strategic moat widens the longer the system operates.

2. Reorganizing Teams and Aligning Culture

Team Structure: In a mid-sized market research practice, a standard project historically required a five-person team: an engagement manager to split deliverables, two analysts to process raw data, one associate to draft findings, and a designer to build client decks. Today, a single senior analyst paired with a proprietary agent can handle the entire pipeline—from data extraction and cleaning to initial drafting. The agent manages mechanical execution, freeing the human expert to focus on defining the underlying business question, auditing findings, and leading stakeholder presentations. Output matches the historical yield of five professionals, while project turnarounds compress from three weeks to four days.

As agents absorb operational volume, human roles shift from hands-on executors to strategic directors and ultimate decision-makers. The middle management layer—task tracking, scheduling, and standard quality checks—maps directly to what agents execute best.

The baseline building block of organizational design is moving away from functional departments and specialized job titles toward lean units: one expert operator managing a dedicated agent cluster. Effective agent orchestration has become a core job requirement, destined to be as universal as standard office productivity software.

Corporate Culture: Adjusting company culture presents a steeper challenge than reorganizing reporting lines. Legacy management defaults to physical presence as a proxy for diligence, evaluating performance on visible hours logged.

Yet an advanced operator might spend two hours in the morning scoping an assignment for an agent fleet and return in the afternoon to audit the output and make strategic adjustments. Though appearing less active throughout the business day, that individual delivers what previously required a full week of labor. Retaining seat-time metrics will penalize an organization's most sophisticated technical operators.

Performance tracking must pivot from inputs to outputs—evaluating what was delivered and its final business impact. Usage data shows that while experienced operators run fully autonomous modes on more than 40% of standard tasks, their rate of manual interventions during unexpected edge cases rises in parallel. Real operational trust requires clear boundaries on when human oversight must step in.

3. Proactive Risk and Security Governance

Recent industry breaches highlight the operational risks inherent in autonomous systems. Summer Yue, Director of AI Alignment at Meta's Superintelligence Lab, experienced a total, unrecoverable deletion of her corporate email archive following an operational misconfiguration within an autonomous agent. Security researchers at Oasis Research uncovered vulnerabilities allowing external web domains to silently hijack local OpenClaw runtimes.

When agents gain native rights to modify local files, send communications, and invoke external APIs, security failures shift from theoretical risks to inevitable operational incidents.

Beyond infrastructure security, management must prepare for two long-term operational risks:

Skill Atrophy: Delegating bulk execution to agents risks eroding foundational judgment and operational instincts among staff who no longer handle frontline tasks. Organizations must maintain deliberate human-in-the-loop training environments.

Compounding System Drift: As autonomous execution chains lengthen, minor misalignments in early stages compound across downstream steps, leading to final outputs that miss baseline business targets. Hard verification checkpoints by human operators remain essential at core business gates.

❓ Frequently Asked Questions

How do AI agents change enterprise management?

They move execution from human-operated software toward autonomous systems, requiring new governance, objective-setting, and control mechanisms.

What should managers redesign first?

Managers should clarify decision rights, permission boundaries, quality assurance, escalation paths, and human accountability.

🎓 Deepen Your Strategic Mastery

Turn these strategic insights into a focused, defensible one-person-company operating system with the Garbo Decodes China mini-MBA.

Explore Garbo Decodes China