Tencent Research Institute July 22, 2026 | Beijing Author: Bo Yang (Tencent Technology)

Late last year, the market largely treated AI as a chatbot—a tool for answering queries, drafting copy, or summarizing documents. Once the Q&A concluded, the AI’s job ended. By the second quarter of 2026, the paradigm shifted. The release of OpenClaw, Codex, and Claude Cowork dragged AI out of the chat window. Agents began autonomously reading files, executing code, formatting spreadsheets, operating software, and integrating directly into internal corporate systems. Assign them a target, and they now deconstruct tasks, call upon tools, and report back upon completion.

OpenClaw's debut in March marked a turning point in market perception. Previously, the consensus viewed agents strictly as coding assistants. With OpenClaw, they began absorbing workloads across diverse sectors. Over the past three months, SOLOMOAT has tracked this "novice" as it integrates into actual production workflows, becoming a new software interface for specialized fields like finance, law, and design.

Corporations recognized the leverage and initiated massive "Tokenmaxxing" campaigns. Yet, after two months of cash burn, this uncapped AI expansion abruptly halted. Executives realized that efficiency gains generated by AI were immediately bottlenecked by human review and decision-making processes. When organizations pivoted to patch these specific operational gaps, latent issues with execution speed and operating costs surfaced. The eight trends summarized in this Q2 report emerge directly from these market frictions. Since Q2 2026, the core operational challenge has transitioned from whether an agent can execute a task, to how to architect an efficient Human-in-the-Loop system while aggressively compressing agent costs.

General Agents Deploy, Resembling an AI OS

Codex, Claude Code (Cowork), and Workbuddy recently transitioned into "General Agents." Why general? OpenAI’s internal reporting shows that between April and May alone, 20% of Codex’s user base consisted of non-programmers, and this segment grew three times faster than its developer cohort. Initially engineered for programming, Codex is now deployed for general utility. Users lacking coding skills require agents. Any repeatable commercial process globally now demands an agent. Powered by Harness (execution frameworks) and Skills (capability modules), Codex can now execute these complex workflows. While not omnipotent, its viable scope is substantial.

Since these agents are general-purpose, the year-long debate over the primary gateway to the AI era can temporarily pause. An entity capable of executing everything naturally becomes the universal interface. Around 2025, the industry bet heavily on the "AI Browser." For two decades, the browser was the definitive internet gateway. Despite the rise of mobile apps, browsers maintained top-tier status. The prevailing logic dictated that to control the AI gateway, one must control the browser, forcing AI to click buttons and fill out web forms like a human. Google launched Project Mariner, OpenAI developed Operator, Perplexity introduced Comet, and rumors even floated regarding a $34.5 billion acquisition of Chrome. Many assumed that if the browser housed human internet activity, it should house agents too.

By SOLOMOAT Editorial Team

One year later, the market has pivoted. Google shut down Mariner in May 2026, merging its assets into the Gemini Agent; Operator was absorbed into the broader ChatGPT Agent. Conversely, Codex, Claude Code, and Cowork interface directly with file systems, terminals, code repositories, data connectors, and local applications—and their usage is scaling significantly faster.

This shift answers the Q1 debate regarding the necessity of the Graphical User Interface (GUI). GUIs are engineered for human cognition; colors and buttons help humans process system architecture. Agents do not need them. Forcing an agent to wait for a webpage to render and simulate a mouse click to hit a button is operationally inefficient. Command Line Interfaces (CLI) and structured data offer a direct, lower-latency route. Under this framework, GUIs survive because humans still need front-end visibility to verify and select data. Within a human-involved loop, the Conversational User Interface (CUI) remains necessary and often more efficient. The browser will not disappear, but it has been downgraded from the "master gateway" to just another tool in the agent's inventory. Data processes in the backend; the UI merely displays outputs for human modification.

Frontier Model Companies Breach Vertical Markets via General Agents

Having secured the gateway, large model developers are pushing into vertical sectors. Anthropic launched Claude Design in April, enabling agents to ingest brand guidelines and codebases to generate design mockups, prototypes, presentations, and marketing collateral. They subsequently released role-specific financial agents—handling valuation audits, general ledger reconciliation, month-end closes, and KYC compliance—before deploying similar architectures in the legal sector.

OpenAI took a different route. It integrated financial, healthcare, scientific, and security capabilities directly into its core models and first-party products, while using Apps, Model Context Protocol (MCP), AgentKit, and Frontier to bridge existing enterprise systems. Despite differing formats and integration depths, both strategies rely on the underlying framework of the General Agent.

Historically, breaking into a new vertical required a model company to execute domain-specific data fine-tuning, engineer custom workflows, and build unique API integrations—a bespoke process for every client. Now, the base agent remains static. To enter finance, developers simply connect the General Agent to financial databases via MCP, equip it with valuation methodologies and compliance protocols (Skills), and allow end-users to customize inputs. To enter law, they swap the finance modules for contract clauses and legal search protocols. The transition from General Agent to Industry Agent only requires swapping domain knowledge, data, and operational rules; the execution environment is fully reusable.

This fundamentally alters the defensive moats of vertical SaaS. With MCP and Harness, assembling a "minimally viable" vertical model only requires purchasing access to specific databases and hiring experts to define operational guidelines. This is a trivial lift for large model companies.

Consequently, the true moat shifts to proprietary enterprise data, permission structures, and historical acceptance records—assets far harder to replicate. For example, tracking whether opposing counsel accepted a specific legal revision, or whether an investment committee overturned a valuation assumption. This operational feedback loop trains the agent for future execution. The companies that capture and apply this feedback earliest secure a structural first-mover advantage.

As long as the foundational model market has not monopolized a specific vertical, early movers retain a data accumulation edge. Thus, vertical capabilities can be mass-assembled for the first time. Yet, as agents absorb higher workloads, human organizational structures often fail to process the resulting output.

Tokenmaxxing: The First Wall Agents Hit

"Tokenmaxxing" was arguably the defining market term of May. Corporations recognized the utility of agents and assumed that allocating more tokens, wider tool access, and longer runtimes would exponentially multiply employee output. Driven by FOMO and inflated expectations of agent autonomy, executives aggressively funded token consumption, treating it as proof of adapting to the AI era so their workforce wouldn't fall behind. Nvidia's Jensen Huang publicly stated that if a $500,000-a-year engineer doesn't burn $250,000 in tokens annually, management should question their AI utilization.

Within three months, this capital burn lost momentum. Amazon’s internal leaderboards incentivized employees to generate useless tasks purely for ranking purposes, forcing management to shut the program down. Uber nearly depleted its entire annual Claude Code budget by April, yet management failed to identify a stable correlation between token consumption and material feature shipping.

The primary friction was cost. Standard queries prompt a model once. Agents executing long-horizon tasks repeatedly query the objective, historical states, tool outputs, and error logs. Token burn for complex operations scales to tens or thousands of times that of a basic query. In late May, Harbin Institute of Technology introduced the "Effective Feedback Compute" metric to calculate the percentage of expended compute that actually alters subsequent steps. For complex tasks, this ratio drops to roughly 10%. The remaining 90% is wasted on re-reading context, failed trial-and-error, and redundant processing loops.

Financial cost is not the only liability. Even if an agent allocates tokens flawlessly, the output often fails to reach final deployment. Generated code still requires human review, testing, integration, and release. Analytical reports require fact-checking and judgment on the final conclusions. Design assets must pass brand, business, and client approvals. When automated processes break, humans must explain the error, execute rollbacks, and absorb the liability.

An MIT study covering over 100,000 GitHub developers found that autonomous coding agents increased code commits by 120%. However, by the project approval phase, this volume compressed to 50%, and only 30% reached production. It mirrors a restaurant tripling its prep cooks' speed while the chefs, plating capacity, and customer demand remain static. The kitchen fills with prepped ingredients, but total sales stall. In economics, substitution theory dictates that a process's total efficiency is capped by its non-automatable segments. AI dramatically accelerates generation, but human review remains sluggish; the agent's leverage is systematically bottlenecked.

Redundant production incinerates further tokens. Because generating a Skill, module, or application is drastically faster, and synchronized organizational tracking is poor, users frequently rewrite existing solutions. Nanyang Technological University analyzed over 20,000 Skills on the market and found approximately 75% were highly redundant; deduplication reduced the total to just over 5,000. Agent-submitted code fixes are routinely rejected simply because another developer already patched the issue. Token consumption spikes, leaving behind a surplus of reinvented wheels.

Demand also fails to track supply. AI can exponentially scale the volume of apps, content, and code, but user attention and willingness to pay remain inelastic. Coding an app is trivial; isolating sustainable, monetizable demand remains exceptionally difficult. Market research indicates app volume surged 40% post-AI, while download metrics flatlined. Tokens cannot fabricate a non-existent market.

The aggressive Tokenmaxxing push failed, but it exposed two critical bottlenecks. Technically, agents waste immense capital; neither latency nor cost is scaling down efficiently. Organizationally, corporate structures for review, judgment, coordination, and liability are missing, meaning upstream efficiency gains cannot be absorbed or effectively monetized. The defining technical shifts of the coming quarter focus specifically on patching these vulnerabilities.

Using Agents to Replace the Human in the Loop

If a single agent executes too slowly, the solution is parallel deployment. If human review is missing, assign a secondary agent to audit the first. Having multiple agents divide labor, cross-examine, and cover blind spots shifts the review burden from humans to the system. This triggered the 2026 Multi-Agent boom.

Currently, the most stable Multi-Agent architecture is the "Orchestrator-Worker" model. A primary agent fragments a task and delegates it to worker agents for parallel execution, then aggregates the outputs. For instance, Claude Research deploys multiple agents to execute independent searches, uses a lead researcher agent to synthesize the data, and assigns a citation agent to verify sources. Kimi Agent Swarm pushes this further, running hundreds of sub-agents concurrently across video, code, and search workflows. This architecture is highly effective for parallelizable tasks, drastically cutting latency and enabling deeper task execution. Kimi reports latency reductions up to 4.5x for specific workloads. Claude Research also demonstrates marked improvements on broad-scope queries.

However, performance gains in this model typically stem from raw compute volume, not genuine collaboration. Anthropic disclosed that multi-agent research systems can consume up to four times the tokens of a standard agent. In select benchmarks, sheer token volume accounts for the majority of performance variance. Today's Multi-Agent system operates like a project manager directing siloed outsourced teams. It parallelizes execution but lacks swarm intelligence. Research indicates that if the central "orchestrator" is removed, allowing agents to negotiate autonomously, human organizational flaws emerge. Some agents default to majority consensus; others succumb to social loafing, assuming "someone else will do it." Testing proves that grouping several agents can actually lower accuracy compared to a single agent holding complete context. Fundamentally, collaboration was never prioritized in their training data. Placing models accustomed to isolated execution into a shared environment does not spontaneously generate collaborative capability. Collaboration is a distinct game theory challenge: one entity's action alters another's environment, and one's judgment shifts another's choices. Therefore, the next evolution for multi-agent systems requires institutional design: protocols for task delegation, information sharing, liability assignment, incentive routing along the workflow, and mechanisms to cull chronically underperforming agents.

A more advanced trajectory is "Self-Improving AI" (RSI). Anthropic highlighted this in their June reporting. They tasked a model with continuously optimizing code used to train smaller models. Driven by advancing base capabilities, the optimization speed scaled from a 3x acceleration on Claude 3 to over 50x on Mythos. The core mechanism mirrors human experimentation in five steps: identify the flaw, hypothesize a fix, structure the test environment and data, validate outputs, retain effective changes, and iterate. Companies like Minimax have integrated these automated loops into post-training, enabling full automation. As long as the target and scoring metrics are absolute—such as "optimize this code execution speed by 50x"—the agent can autonomously isolate bugs, patch code, run tests, and lock in viable upgrades continuously. It executes this faster and more relentlessly than human engineers.

Yet, it still lacks directional intuition. When evaluating "which research path holds commercial value" or "is this metric misleading," the agent falters. These decisions require navigating massive search spaces using Market Clarity and judgment. In Anthropic's trials involving complex directional decisions where humans had previously failed, the model outperformed humans only 20% of the time. When humans excel at a task, the model's competitive edge vanishes.

Another concept that surged in June is Loop Engineering, which structures these cycles into persistent, long-running operations. Agents no longer wait for manual prompts; they proactively hunt for tasks, execute, validate, log feedback, and determine the next operational cycle autonomously.

While these represent technical advancements, the underlying economics still require correction.

The Unviable Economic Math

Historically, the GPU dominated the AI compute narrative. Training large models demands massive matrix calculations, which GPUs execute perfectly. CPUs were relegated to basic program initiation and data prep, maintaining a low profile. Agent workloads operate differently. They are not single-pass transactions; they continuously pivot between reasoning, API calls, and waiting for external outputs. Significant time is burned outside of model inference: commanding browsers, managing file systems, and handling timeouts. Any bottleneck in this chain forces the high-capex GPU to sit idle.

A late 2025 paper, "A CPU-Centric Perspective on Agentic AI," evaluated five agent categories, finding tool processing accounted for up to 90.6% of total task latency, with dynamic CPU power consumption hitting 44% of total system draw. Co-optimizing CPU and GPU task scheduling improved median latency for specific workloads by over 2x. Consequently, the CPU is reclaiming its position at the core of compute architecture. While the GPU handles inference, the CPU maintains the concurrent environment, manages task queues and tool calls, while KV cache and memory preserve each agent’s sandbox, logs, and intermediate states. Networking protocols route data between chips and servers. Congestion anywhere in this pipeline leaves the GPU stranded.

Capital markets are actively pricing in this pivot. AMD reported Q1 2026 data center revenue of approximately $5.8 billion, a 57% year-over-year increase, with server CPU revenue jumping over 50%. The company doubled its 2030 server CPU TAM forecast to $120 billion, explicitly citing the scheduling, data movement, and execution demands driven by agents.

Beyond latency, model pricing is bifurcating. By Q2 2026, the cheapest mainstream models cost mere cents per million input tokens, while premium frontier models command tens of dollars. Two years ago, this spread was roughly 30x; today, it approaches 600x. Commoditized models capture increasing token volume, while premium models retain mission-critical tasks and majority revenue. A vast array of processes—repository scanning, data classification, extraction, and log formatting—only require cheap models. Only complex refactoring, security audits, legal assessments, and critical post-failure decisions justify premium pricing.

Currently, this pricing stratification is not automated. ChatGPT introduced intelligent routing to allocate models based on task complexity, but poor execution limited adoption. However, in June, Sakana released Fugu, which achieved exceptional results by abandoning rule-based routing and training a dedicated Agent specifically for the routing function. It ingests the task, dynamically erects operational scaffolding, and deploys a temporary team of models varying in price and capability. The premium model executes only critical nodes, outsourcing the rest to cheaper variants. This achieves top-tier output at half the cost.

Software engineering will likely perfect this architecture first due to its immediate feedback loops. Repositories, code diffs, testing suites, linters, CI pipelines, and error logs instantly signal to the router whether the previous allocation succeeded. Cheap models read code and generate tests; mid-tier models execute standard patches; premium models handle core refactoring and security reviews, leaving final validation to automated tools. Products like Codex and Claude Code will inevitably adopt automated routing, evolving into engineering organizations that manage model teams. They will deploy budgets across a portfolio of models rather than relying on a single engine end-to-end.

Yet, the friction remains. As agents scale, they are structurally altering employment, security, information integrity, and human cognition—and not necessarily for the better.

The Contracting Reality

Previously, the negative externalities of agents existed primarily in theoretical models. By Q2 2026, these impacts have materialized in real-world labor markets and product pipelines.

Labor markets absorb the initial shock. AI’s threat to employment does not immediately register as a macroeconomic spike in unemployment. It alters hiring pipelines, daily task allocation, and corporate budgeting. Customer service, IT support, ticketing, internal Q&A, baseline operations, and segments of data Analysis, DevOps, and project management are already being absorbed by agents. Other roles aren't directly automated but are eliminated by AI capital expenditure. GPUs, data centers, and model teams require massive funding, forcing enterprises to cannibalize other departmental budgets. Surviving personnel pivot to managing agents, monitoring workflows, and executing complex decisions; the job title remains, but the core function is entirely rewritten.

Entry-level positions face extreme pressure. Data retrieval, summarization, basic coding, and client data formatting are trivial tasks for agents, yet these are exactly the repetitions junior staff need to master industry mechanics. Humans build business acumen, decode codebases, and learn organizational judgment through this baseline work. A May 2026 paper, "Generative AI and the Reorganization of Labor Demand," analyzed job postings and found that since 2023, the share of highly automatable job tasks dropped by approximately 10%. Half of this decline resulted from companies eliminating these roles entirely; 40% came from rewriting existing roles to emphasize tasks resistant to AI. Enterprises demand "experienced junior talent" while simultaneously automating the very tasks that generate that experience. The career entry funnel is structurally narrowing—a shift far more permanent than a standard cyclical layoff.

Beyond labor, security liabilities are migrating from academic papers to product launches. Anthropic’s Mythos demonstrated severe cybersecurity capabilities, successfully identifying and verifying high-severity zero-day exploits. Management refused to release it as a standard model. Yet, after imposing guardrails on Falbe 5, developers and researchers complained the constraints rendered it operationally useless. OpenAI’s frontier models are also pivoting to restricted access frameworks. Broad open-source access lowers the barrier for cyberattacks and biosecurity threats. Excessive restriction suffocates legitimate research and commercial value. Determining who accesses premium models and defining "trusted partners" is now heavily dictated by government regulation and geopolitics. Model developers must not only engineer superior capabilities but definitively prove they can deploy them safely. Security is now the primary choke point for frontier models.

The internet is simultaneously generating a novel information loop. While AI-generated content has not swallowed the entire web, contamination within AI search protocols is escalating. Graphite’s analysis indicates that between early 2024 and 2026, AI-generated material stabilized at 48% to 50% of total internet content. However, within ChatGPT’s citation sources, the ratio of AI-generated content jumped from 38.9% to 42.7% in just six months. The model's historical outputs are being recycled as empirical evidence for future queries. As SEO and GEO operators exploit this bias, search ecosystems risk degrading into closed echo chambers of machine-to-machine citations. Eventually, verifiable human sourcing—content generated by individuals who visited the site, interviewed the subject, edited the text, and attached their real name—will trade at a massive premium.

Beyond macroeconomic and structural impacts, agents trigger individual "cognitive surrender." When a problem surfaces, users instinctively ping ChatGPT before applying critical thought. Before an opinion forms, the AI has generated the outline. Hesitation, trial-and-error, skepticism, and verification are bypassed by a frictionless, generated answer; the user forfeits the cognitive process of forming a judgment. Wharton researchers tracked 1,372 participants across 10,000 reasoning tasks. When permitted, participants utilized AI on over 50% of the prompts. Even when the AI supplied demonstrably false data, roughly 80% of users deferred to it. Anthropic’s developer study revealed that while AI-assisted cohorts executed novel tasks faster, their performance on subsequent closed-book tests dropped by an average of 17%. The execution is outsourced; the capability is not retained. The report carries your signature, the code sits under your commit history, and the strategy is pitched by you. You retain the liability, but the critical thinking behind the deliverable is increasingly outsourced.

Taylorism reduced factory workers to physical appendages on an assembly line; AI threatens to reduce knowledge workers to rubber stamps on a cognitive assembly line.

Therefore, future AI architectures must intentionally engineer friction into the loop. Mandate humans to explain and verify before executing high-risk operations; force users to document their initial judgment before revealing the AI’s conclusion; present contradictory evidence rather than seamlessly automating every step.

This degrades raw speed, but it preserves the operational space for critical thought and human capability scaling.

Final Note

The eight market trends from Q2 2026 form a contiguous narrative. General Agents hijacked the software interface, allowing frontier model companies to breach vertical markets. Tokenmaxxing rapidly stalled, exposing the structural bottlenecks of human review and system compute costs. Multi-Agent systems and RSI are attempting to automate that review and iteration process, while CPU dominance, advanced memory architecture, and model routing attempt to drive down execution costs. As these technical blockers clear, the secondary impacts—labor contraction, security liabilities, information pollution, and cognitive surrender—are surfacing.

Moving forward, corporations must focus on accumulating structural AI assets, not just tracking token burn. Every execution cycle must generate reusable workflows, evaluation metrics, permission architectures, organizational memory, and empirical feedback. Otherwise, when the capital budget runs dry, the enterprise is left with nothing but an API invoice. At the individual level, competing with AI on generation speed is a strategic error. Machines can rapidly generate answers, but humans must retain the Market Clarity to determine if the problem is worth solving, project the commercial consequences of the answer, and bear the ultimate liability.

Frequently Asked Questions

Why do AI agents struggle outside controlled settings?

Real environments contain incomplete information, changing constraints, exceptions, physical dependencies, and responsibility that cannot be assumed away.

How should founders use agents?

Use agents inside carefully bounded workflows, create escalation paths, and learn from failures before expanding autonomy.