Executive Summary
Tencent Research Institute’s flagship 2026 report, Co-Evolution: 2026 AI Ten Major Trends, delivers the most comprehensive roadmap yet for understanding how artificial intelligence is reshaping every aspect of technology, business, and society. Co-produced with Tencent Cloud Intelligence and Tencent Youtu, the report maps out three tectonic shifts: Evolution (how models learn, perceive, and discover), Landing (how AI becomes reliable infrastructure), and Restructuring (how AI rewires organizations and society).
This article distills the report’s 10 trends into an actionable framework, complete with data tables, competitive intelligence, and forward-looking projections. Whether you are a C-suite executive, product leader, engineer, or investor, this analysis delivers the signal you need to navigate 2026–2027.
Part I: Evolution — How AI Learns, Perceives, and Discovers
Trend 1: Online Evolution — Learning Doesn’t Stop at Training
The dominant paradigm of 2024 was Scaling (more data, more parameters, more GPUs). 2025 was the year of Reasoning (RLVR-driven emergent chain-of-thought). 2026 marks the arrival of Context Learning — the recognition that models must continue learning after deployment, adapting to real-world usage in real time.
The Four-Stage Progression:
| Stage | Year | Core Mechanism | Key Finding |
|---|---|---|---|
| Scaling | 2024 | Larger models, more data, more compute | Models kept getting smarter, but marginal returns began declining |
| Reasoning | 2025 | RLVR (Reinforcement Learning with Verifiable Rewards) | Reasoning emerged without supervised fine-tuning; code and math were breakthrough domains |
| Context Learning | 2026 | In-situ learning from user context and feedback | Frontier models solve only ~17% of tasks requiring context-based learning (CL-bench) |
| Memory Consolidation | Future | Cross-session knowledge persistence | Tencent Hunyuan’s Hy-Memory scores 85.2 on LongMemEval; 3-4x higher information density |
The Key Insight: The strongest model is no longer necessarily the largest. Tencent’s Chief AI Scientist Yao Shunyu led the CL-bench experiment, which revealed a sobering truth: even the most advanced frontier models are surprisingly poor at learning from context. When tested on entirely fictional content (made-up interstellar laws, fake SDKs), the best models achieved a task completion rate of only ~17%. They defaulted to pre-trained knowledge rather than integrating new contextual information.
RLVR’s Concentric Expansion: Reinforcement Learning with Verifiable Rewards is spreading beyond code and math into scientific frontiers:
| Domain | Status | Key Signal |
|---|---|---|
| Code / Math | Near ceiling | Power-law curves beginning to flatten |
| Virology | Breakthrough | VirBench: 99.7% accuracy with deterministic retrieval layer |
| Biology, Chemistry, Clinical Medicine | Accelerating | Infrastructure barriers being cleared |
| Materials Science, Superconductors | Early stage | Far from saturation |
| Open-ended scenarios (ARC-AGI-3) | <1% for all models | Marks potential boundary of RLVR paradigm; humans score 100% |
AI Accelerating AI Research: The Self-Improvement Engine Shifts Gears
Three tech giants have converged on a critical insight:
| Organization | Key Disclosure | Timeline Signal |
|---|---|---|
| Anthropic | >80% of production code written by Claude; training code optimization jumped from 3x to 52x in one year; Claude agents recovered 97% of performance gap in 800 hours (human researchers: 23% in one week) | Recursive self-improvement “may arrive sooner than most institutions are prepared for” |
| OpenAI | “Intern-level researcher” by September 2026; fully autonomous research AI by March 2028 | Core bottleneck: sustained autonomy, not intelligence level |
| Google DeepMind | AlphaEvolve recovered 0.7% of Google’s global compute resources; FlashAttention accelerated 32.5%; solved 300-year-old kissing number problem in 11 dimensions | Bypasses human research methods; brute-force searches in impossible-to-enumerate design spaces |
The Autonomy Gap: AI reliably completes tasks for 16 hours in 2026 (up from 4 minutes in 2024), doubling every 4 months. But autonomous research requires weeks of continuous operation. Human researchers’ comparative advantage remains “research taste” — knowing which problems are worth solving, when to trust results, and when to change direction.
12-Month Prediction Checkpoints: (1) Will RL-driven reasoning produce scaled commercial products in ≥3 non-code domains? (2) Will practical-use models become the default choice in most enterprise scenarios? (3) Will recursive self-improvement from OpenAI and Anthropic arrive earlier than expected?
Trend 2: Multimodal Cognition — From Renderer to Creator
2026’s multimodal AI is undergoing a fundamental role shift. The industry spent two years optimizing for “more realistic” output. Now the game has changed: generative systems are exhibiting intent understanding, step planning, and autonomous iteration — generation itself has become a multi-step decision process.
The Three-Layer Cognitive Ascent:
| Layer | Capability | 2026 Breakthrough | Significance |
|---|---|---|---|
| Controllability | Stable, reproducible generation | ByteDance Seedance 2.0: ~90% usable rate for 15-second video (2025: ~20%) | Crosses threshold for production workflows |
| Editability | Semantic-level local modification while preserving continuity | Selective reconstruction of visual content | Shifts from “one-click output” to “generate-feedback-modify” iterative collaboration |
| Agentification | Intent-driven autonomous planning and multi-step creative decision-making | Kling 3.0: smart storyboarding as native feature; GPT Image 2: native multimodal architecture | Interaction paradigm shifts from “describe the image” to “orchestrate the narrative” |
Native Multimodal Architecture: Stitching the Modality Gap
The “language model schedules, vision model executes” split architecture has a fundamental flaw: intent gets simplified and distorted during modality translation. Tencent Hunyuan team’s experiments proved that even GPT Image, when handling metaphor transfer, only performs literal substitution — it cannot abstract relational logic.
The next-generation architecture encodes images, text, and audio into a shared latent space, enabling cross-modal attention at the token level. This enables emergent visual reasoning: a model can infer social trends from a chart, read cultural metaphors from an architectural photograph, or proactively complete missing logical segments in a video — all without text prompts.
World Models: Toward Action
Li Feifei’s 2026 classification framework decomposes world models into three levels:
| Level | Examples | Current Status |
|---|---|---|
| Renderer | World Labs’ Marble | Operational |
| Simulator | Tencent Hunyuan HY-World 2.0 | Advancing |
| Planner | NVIDIA Cosmos 3, DeepMind Genie 3 | Approaching |
A critical experimental finding: pure video generation models can perfectly simulate physical motion within training distribution, but error increases by an order of magnitude when encountering unseen condition combinations. This has revived interest in non-generative approaches: Meta’s V-JEPA 2-AC achieves zero-shot robot control with just 62 hours of unlabeled video, running ~15x faster than generative methods. Yann LeCun’s long-held position is being validated: understanding the world does not require pixel-perfect reconstruction.
Multimodal Memory: From One-Shot Generation to AI Companions with “Visual Experience”
The next frontier is persistent multimodal memory at three layers:
- Situational memory: Remembering the user’s home, office, children’s growth — maintaining spatiotemporal consistency
- Preference memory: Learning color palettes, composition habits, and narrative rhythms from repeated edits and selections
- Emotional memory: Accumulating “shared history” from co-created designs and restored photographs
This raises acute privacy challenges: visual memory is far more intimate than text memory. The right to be forgotten and value alignment mechanisms for the visual domain must be built in parallel with capability development.
Trend 3: Scientific Intelligence — The Trinity at Scale
AI for Science (AI4S) entered 2026 with industrial maturity accelerating rapidly across three converging waves: technology system construction, industrial capital accumulation, and infrastructure rebuilding.
The Trinity Framework: Foundation Model + Research Agent + Autonomous Laboratory
Foundation Models as Scientific Operating Systems:
| Organization | Initiative | Scope |
|---|---|---|
| Gemini for Science AI toolset + Co-Scientist agent | Unified workflow: literature insight → hypothesis generation → computational discovery | |
| Anthropic | Mythos 5 + Claude Cowork + ~$400M acquisition of Coefficient Bio | End-to-end drug discovery from literature to clinical treatment |
| Shanghai AI Lab | Intern-S1-Pro (trillion-parameter scientific multimodal model) + “AGI4S Everest Plan” | Open-source ecosystem across upstream and downstream |
| Chinese Academy of Sciences | Panshi Scientific Foundation Model | Covers 8 disciplines, 50+ institutes, 100+ research scenarios |
Dedicated Scientific Models: A Deep Matrix Across Disciplines:
| Domain | Key Models & Capabilities |
|---|---|
| Life Sciences | Alpha series (protein sequence → function → generation → regulation); GPT Rosalind series (GPT-5.5-based); Amazon Bio Discovery (antibody development) |
| Materials Science | DPA4 (next-gen atomic model architecture); MatterSim-MT (Microsoft, multi-property materials simulation) |
| Physics & Earth Science | Shanghai AI Lab & CAS models (top-ranked in benchmarks); AlphaEarth (satellite remote sensing) |
Autonomous Labs Approaching Scale:
Self-Driving Labs (SDL) follow a “Design → Execute → Characterize → Learn → Redesign” loop with decision-making partially delegated to AI — a key differentiator from traditional high-throughput automation.
| SDL Layer | Function | Technology |
|---|---|---|
| Experiment Planning | Decide next experiment | Active learning, Bayesian optimization, LLM agents |
| Orchestration & Scheduling | Unified device coordination | UniLabOS, ChemOS, MADSci |
| Robotic Execution | Modular experiment running | Liquid handlers, robotic arms, mobile robots |
| Online Characterization | Real-time monitoring | HPLC, NMR, XRD, spectroscopy |
| Data & Knowledge | Reproducibility & portability | Metadata, knowledge graphs, xDL protocol |
Commercialization milestones:
- XtalPi (China): orchestrates hundreds of workstations; 7×24 R&D; produces 50,000+ reaction yield data points and 300,000 process data points monthly; achieved first annual profit in 2025
- Lila Sciences (US): $550M cumulative funding; signed 235,500 sq ft lab lease in Greater Boston
- Berkeley Lab’s A-Lab: >2 new materials discovered per day
Four Application Domains at Tipping Points:
| Domain | Stage | Key Signal |
|---|---|---|
| Biomedicine | Pre-commercial | ~170 AI-discovered/designed drug projects in clinical development; multiple in Phase III; Insilico Medicine: 13 IND-approved candidates |
| Materials Science | Early commercialization | Startups (Periodic Labs, CuspAI) centered on materials; shorter experimental feedback loops than pharma |
| Meteorology | Deepest AI4S penetration | AI models saving >90% compute vs traditional numerical models; 0-15 day medium-range forecasting as primary pilot |
| Mathematics | Inflection point | GPT-5 applied to Erdős problems; AlphaEvolve advancing theoretical CS; Olympiad gold-medal performance; independent problem definition remains distant |
The π-Shaped Talent War: The bottleneck is not AI researchers learning science or scientists learning AI — it is the scarcity of native “π-shaped talent” who deeply integrate both. Global competition has intensified: Anthropic’s AI for Science Program (free API credits + STEM Fellows Program), Google’s $30M AI4S Impact Challenge, OpenAI Residency (6-month cross-disciplinary training), and aggressive university-level interdisciplinary programs at Tsinghua, Peking, and Shanghai Jiao Tong universities.
Trust Crisis: AI-generated fake references are exploding. A Lancet study analyzing 2.5 million biomedical papers and 126 million citations found fictitious references rising from ~4.0 per 10,000 papers in 2023 to 56.9 in early 2026 — a >12x increase. Nature estimates over 110,000 academic publications globally may contain at least one invalid citation in 2025.
Part II: Landing — From Usable to Reliable Infrastructure
Trend 4: Harness Engineering — Taming the Wild Horse
In early 2026, within a single month, both Anthropic (“Harness Design for Long-Running Applications”) and Cursor (“Towards Self-Driving Codebases”) published manifestos on the same topic: how to build an environment that lets models autonomously complete tasks. The emerging discipline is Harness Engineering.
The Evolution of Engineering’s Frontier:
| Era | Focus | Core Question |
|---|---|---|
| 2022–24: Prompt Engineering | How to ask the model | Getting instructions right |
| 2025: Context Engineering (Karpathy’s reframing) | What to put in the context window | Curating the model’s working memory |
| 2026: Harness Engineering | What rules the model runs under | Memory, tools, planning, recovery, multi-agent coordination — the entire runtime container |
Why Now? The marginal returns from scaling pre-training are declining, and the data wall is approaching. Meanwhile, inference-time compute has emerged as a new force multiplier. Microsoft Research (2025) demonstrated that with fixed model parameters, techniques like multi-sampling, tree search, and self-verification can significantly boost complex task performance. Orchestrating these inference steps is exactly what Harness does.
Three Independent Growth Vectors:
| Vector | Improvement Mechanism | Example |
|---|---|---|
| Model Intelligence | Stronger reasoning | Better base models |
| Context | Cleaner working memory | Curated information in context window |
| Harness | More reliable runtime framework | Memory management, tool selection, error recovery, multi-agent orchestration |
The Terminal State: Harness will eventually be internalized by models themselves. Today’s externally built memory management, error recovery, tool selection, and multi-step planning will become native model capabilities. Every new model generation absorbs the Harness capabilities that the previous generation required externally. The current window — where Harness is visible, named, and being engineered — is temporary. Once internalized, it becomes invisible infrastructure, like a saddle: riders don’t think about how it fits the horse, they just know the ride is more stable.
ADPS (Agent Design Patterns Society) has catalogued 28 named design patterns across two axes, with Manning publishing Designing AI Agents as a selection handbook. This represents the industry’s push to turn Harness into an engineering discipline.
The Shift in Work: The focus moves from “commanding every step of AI” to “designing how AI works.” Instead of pursuing step-by-step certainty, practitioners pursue outcome certainty: define acceptance criteria clearly, let a stateful Harness manage the process, and let AI self-discover and self-correct until reaching the verification standard. Human effort shifts from micro-monitoring to defining endpoints and designing environments.
Evaluation as the Control Layer: Evaluation systems are being moved to the front of the architecture: they define task acceptance criteria, record agent execution trajectories, judge whether each step deviates, convert errors into traceable evidence, and feed that evidence back into the next run. Without evaluation, Harness is just tools plugged into a model. With evaluation, it becomes a self-verifying, experience-accumulating engineering system. Evaluation data becomes the feedstock for model learning — this is the concrete path for “Harness being internalized by models.”
Trend 5: Universal Building — When Everyone Becomes a Builder
AI coding was the proving ground for the Agent wave. By 2026, the debate about whether AI can independently deliver software is effectively settled.
The Capability Leap on SWE-bench Verified:
| Model | Score | Date | Delta |
|---|---|---|---|
| Claude 3.5 Sonnet | 49% | Jun 2024 | — |
| Claude Opus 4.5 | 80.9% | Nov 2025 | +31.9pp |
| Claude Opus 4.7 | 87.6% | Apr 2026 | +6.7pp |
| Total: | 22 months | +38.6pp |
The qualitative shift happened on two dimensions: autonomy (continuous operation from single-line completion to hours-long multi-file, multi-repository engineering) and self-verification (models performing formal verification before writing system-level code — shifting from “write then check” to “prove then write”).
Production Dominance Data:
| Organization | Metric | Value |
|---|---|---|
| Anthropic Claude Code team | Code written by Claude Code | 95% |
| Anthropic Claude Cowork | Entire product built by AI | 1.5 weeks |
| Tencent | Engineers using CodeBuddy | >90% |
| Tencent | Average coding time reduction | >40% |
| YC W25 batch | Startups with >95% AI-generated codebase | 25% |
| Anthropic | % of developers refusing “no-AI” work conditions | 30–50% |
Commercial Velocity:
| Product | Milestone | Timeline |
|---|---|---|
| Claude Code | $0 → $2.5B revenue | 9 months |
| Cursor | Valuation discussion range | $29.3B → $50B |
| Codex | Weekly active users | 3M → 5M (4 months) |
| Fortune 500 | Formal paying customers of AI coding startups (a16z survey) | 29% |
The Democratization of Building:
Platforms like Bolt, Lovable, and v0 have made “natural language → running software” accessible to non-programmers. Bolt reports 60-70% of its users are designers, students, fitness coaches, salespeople, and teachers — not coders. Epic Games: over half of Claude Code internal usage comes from non-developers. Block: thousands of non-engineers are building their own tools connecting to business systems.
The “Prototype Wall”: AI has compressed “zero to prototype” cost to near zero, but “prototype to operational” remains a chasm. Data reveals the crisis:
- GitClear analysis of 211 million lines of code: AI intervention → 30–40% increase in technical debt, code duplication rate doubled
- Veracode testing: 45% of AI-generated code tasks introduce known security vulnerabilities
- Lovable-style platforms: common user pattern of “Week 1 excitement → Week 3 anxiety → Month 2 abandonment”
The Spillover: Coding Agents Become Universal Work Agents
In the first half of 2026, leading coding products nearly simultaneously evolved into general-purpose Work Agents:
| Product | Parent | Launch | Positioning |
|---|---|---|---|
| Claude Cowork | Anthropic | Feb 2026 | “Claude Code for non-coders” |
| WorkBuddy | Tencent Cloud | Mar 2026 | Derived from CodeBuddy |
| Codex (expanded) | OpenAI | Jun 2026 | 6 role-based plugins; data analysis tasks +110% WoW |
| Kimi Work | Moonshot AI | 2026 | Upgraded from Kimi Code |
| TRAE Work | ByteDance | 2026 | Covering office + development |
The Spillover Formula: AI coding was the ideal proving ground because it satisfies three conditions simultaneously — tasks definable in natural language, outputs instantly verifiable (compile/run), and workflows already standardized (version control, code review). Any domain meeting these three criteria is next in line: design, law, finance, consulting, healthcare clinical pathways, education personalization, government standardized approvals, industrial maintenance.
Trend 6: Trusted Execution — Building Trust into the Pipeline
As Agents move from intelligent assistants to autonomous executors capable of calling tools, connecting systems, and completing multi-step tasks on behalf of humans, security risk amplifies proportionally.
Two Wake-Up Calls in 2025:
| Incident | What Happened | Implication |
|---|---|---|
| EchoLeak (CVE-2025-32711) | Zero-click prompt injection in Microsoft 365 Copilot; hidden instructions in emails caused AI to silently exfiltrate enterprise data during auto-summarization | Traditional security models fail at machine speed |
| First Near-Autonomous AI Cyberattack | Attacker claimed to be conducting authorized penetration testing; Agent autonomously completed ~80-90% of attack actions | Model safety guardrails can be socially engineered |
MCP Security Crisis: By April 2026, OX Security disclosed architecture-level vulnerabilities in the Model Context Protocol (MCP), affecting all official SDKs (Python, TypeScript, Java, Rust), with ~200,000 potentially affected instances across 32,000+ code repositories. Over 40 CVEs were filed against the MCP ecosystem in January–April 2026 alone.
The Five-Layer Trust Framework:
| Layer | Core Function | Key Standards & Technologies (2026) |
|---|---|---|
| Application Layer | Real-time interception, continuous security testing | Guardrails AI (<50ms), OWASP ASI Top 10, CI/CD-integrated red-teaming |
| Standardization Layer | Interoperability baseline | GB/Z 185-2026 (China AIP national standard), NIST COSAIS, ISO 42001 AIMS |
| Data Flow Layer | Source verification, semantic-level traceability | OpenTelemetry semantic tracing, RAG, ASL trusted intent verification |
| Identity Layer | Unique identification, mutual authentication | AIC Agent Identity Code (China), SPIFFE/SPIRE, mTLS-A, Signed Agent Cards (JWS) |
| Environment Layer | Compute base, zero-trust isolation | MXC Execution Container (kernel-level sandbox), Zero-Trust MCP Gateway |
The Identity Imperative: Agents can no longer be treated as anonymous scripts. Two 2026 incidents underscore this: Step Finance lost ~$30M when an attacker compromised an executive’s device and induced its AI trading agent to transfer assets; Mexican government systems were breached with AI agents impersonated to autonomously execute ~75% of remote commands, exposing ~195 million taxpayer records.
Regulatory Convergence: China’s CAC, NDRC, and MIIT jointly issued the Implementation Opinions on Regulating Agent Applications and Promoting Innovative Development (May 2026), shifting regulatory focus from “managing models” to “managing actions.” The EU AI Act requires AI to disclose identity at the point of interaction. TC260 published Agent Security Standardization Research proposing verifiable digital identity cards for every agent.
The operational principle: clear identity at entry → traceable chain in the middle → minimized permissions at exit. Every autonomous action must be constrained within an end-to-end trusted pipeline.
Part III: Restructuring — When AI Becomes a Participant in Society
Trend 7: Intelligence as a Service — Token Retreats, Intelligence Takes Center Stage
Intelligence has historically been the hardest factor of production to marketize — it manifests as human judgment, experience, and situational handling, inseparable from specific people and contexts. Large models are loosening this constraint.
The Core Tension: Tokens record consumption but cannot measure value. The same 10,000 tokens might polish a single sentence or identify critical risk in a multi-million-dollar contract. Yale Cowles Foundation economists call tokens a “contractible unit of measurement” — quantity can be precisely recorded and written into contracts, but value cannot be captured in the same way.
Market Reality from OpenRouter Data (March 2026):
| Price Range (per million tokens) | Share of Total Call Volume |
|---|---|
| Free – $1 | ~43% |
| $1 – $2 | ~14% |
| $2 – $3 | ~10% |
| $3 – $4 | ~10% |
| Total <$4 | ~76% |
| $4+ (across all higher brackets) | ~24% (highly fragmented) |
76% of real-world developer spending concentrates below $4 per million tokens. Higher-priced models, despite superior capability, see rapidly declining call volume. Token price wars are fierce at the bottom; the real competition has moved up.
The Commercialization Spectrum: From Selling Resources to Selling Job Roles:
| Pricing Model | What’s Sold | Platform Risk | Buyer Comparison Point | Maturity |
|---|---|---|---|---|
| Pay-per-use | API calls (raw resources) | Extremely low | $/M tokens | Commodity |
| Subscription | Access optionality | Low | Monthly fee | Consumer-friendly |
| Credit-based | Consumption units | Low | Credits | Consumer-friendly |
| Per-workflow | Process closure | Medium | Per action | Process-oriented |
| Per-outcome | Verifiable completion | Medium-High | Per result | Outcome-oriented |
| Digital employee | Job role capability | High | Full cost of human role | Most valuable |
Three Profitability Pathways (ascending margin ceiling):
| Pathway | Representatives | Margin Logic |
|---|---|---|
| API Pricing Power | Zhipu (raised prices during industry price war, volume increased); Claude (premium pricing, sustained demand) | Capability differentiation + ecosystem lock-in; not all APIs are commoditized |
| Subscription Scale | OpenAI (ChatGPT Plus + Enterprise seats) | Stable cash flow + strong entry point; ceiling: cost mismatch from multi-modal/long-context/Agent inference |
| Agent Productization & Role-based Pricing | Claude Code, Codex, AutoClaw | Token costs managed internally (caching, routing, model tiering); revenue anchored to job role value → software-level margins possible |
The Token Formalism Risk: A cautionary signal from 2026 — Meta internally leaked a leaderboard ranking employees by Token consumption. Some employees deliberately ran useless agent tasks to appear “AI-native.” When consumption metrics become capability metrics, tokens transform from production tools into performance tokens. Governance must shift from minimizing token usage to managing intelligence budgets.
Trend 8: The Agentic Internet — When Agents Become the New Citizens of the Web
Mobile internet took 15 years to connect 5 billion people. Now a new type of user is flooding the network — and it is not human. AI Agents do not scroll through short videos, view advertisements, or need polished UI. They are calling services, completing transactions, and collaborating with other Agents at machine speed.
The Agentic Internet rewrites three foundational assumptions:
| Dimension | Old Paradigm (Attention Economy) | New Paradigm (Outcome Economy) |
|---|---|---|
| Metrics | DAU, time-on-platform | TCR (Task Completion Rate) |
| Pricing | Per-impression, CPM | Per-outcome, per-savings-share |
| Competition | Traffic distribution (SEO, ads, feeds) | Capability orchestration (API stability, verifiability, call success rate) |
| Service Provider Competitiveness | Brand exposure, page conversion | Interface reliability, result verifiability, historical call success rate |
Multi-Agent Collaboration Architecture:
The dominant architecture in 2026 is a Master Agent that decomposes complex objectives and orchestrates multiple specialized sub-agents in parallel:
code复制
Human (Supervision, Decision-making)
│
▼
Master Agent (Overall Control, Orchestration)
│
├── Research Agent (Information retrieval, analysis)
├── Execution Agent (API calling, implementation)
├── Audit Agent (Cross-verification, error correction)
└── Creative Agent (Generation, divergence)
Evolution Pace:
- 2024: “Can AI complete a task?”
- 2025: “Can AI complete an entire workflow?”
- 2026: “Can AI replace a corporate functional department?”
McKinsey, Deloitte, and Gartner’s 2026 technology trend reports all independently identified multi-agent orchestration as a core direction. Task Completion Rate (TCR) is replacing DAU and time-on-platform as the North Star metric for next-generation internet products.
The New Capability Flywheel: More callable capabilities → more complex tasks completable → more users willing to delegate to Agents → more service providers integrating → platform orchestration capability strengthens → more callable capabilities.
Vertical Agent Penetration: IDC data shows >45% of global enterprises have deployed AI Agents with autonomous decision-making capability in core business processes, with an average ROI of 171%. The penetration follows a three-phase model:
| Phase | Description | Examples |
|---|---|---|
| Single-Point Automation | High-frequency, rule-based, repetitive tasks | Customer service responses, data entry, document generation |
| Process Orchestration | End-to-end workflow spanning multiple steps | Insurance claims: material collection → clause comparison → risk assessment → payout recommendation (compressed from days to hours) |
| Organizational Restructuring | Department-level capability replacement | Functions handled by Agent clusters + minimal human supervisors |
Sectors by Adoption Speed:
- Fastest: Finance, customer service, legal, HR (text-heavy, rule-based, process-driven)
- Moderate: Manufacturing, healthcare, scientific research (requiring physical operations or high-precision judgment)
Vertical Agent competitive moats are not in the underlying model — they lie in three layers: (1) depth of accumulated industry data and knowledge, (2) depth of business system integration (ERP, CRM, HR, development environments), and (3) depth of workflow understanding (knowing not just the steps, but the critical risk points specific to that industry).
Trend 9: Liquid Organization — From Solid Structures to Liquid Orchestration
February 2026: Block announced 4,000 layoffs (~40% of workforce). One month later, founder Jack Dorsey published “From Hierarchy to Intelligence,” revealing the underlying logic: replace middle management’s information-routing function with AI, restructuring the company around intelligent orchestration.
The Acceleration of Organizational Restructuring:
| Company | Action | Scale |
|---|---|---|
| Block | Layoffs + restructuring | 4,000 employees (~40%) |
| Meta | Layoffs + hiring freeze + AI department transfers | 8,000 laid off; 6,000 positions frozen; 7,000 transferred to 4 new AI departments |
| Anthropic | Minimal growth team for massive revenue | ~40 people in growth team; $14B annualized revenue |
| Pieter Levels | Solo-operated multiple products | Multi-million-dollar annual revenue as one-person company |
| Cursor | Minimal team for major revenue | ~60 employees at $300M ARR stage (traditional: 500–1,000) |
HBR 2026 finding: Companies are cutting jobs for AI’s potential, not AI’s proven results. Many roles have not yet been truly replaced by Agents, but organizations are already restructuring around the anticipated replacement.
From Solid to Liquid:
| Dimension | Solid Organization | Liquid Organization |
|---|---|---|
| Basic Unit | Human | Human + AI Agent |
| Structure | Rigid pyramid | Dynamic, flat network |
| Decision-making | Top-down centralized | Autonomous, close to the front line |
| Work Method | Functional division, accountable for components | Task-based division, accountable for outcomes |
| Person-Organization Relationship | Employment-based | Partnership-based, diversified |
Silicon-Carbon Mixed Teams: McKinsey’s 2025 archetype: 2–5 humans supervising 50–100 specialized Agents, supporting end-to-end processes like client onboarding, product launch, and financial close. Shopify’s CEO mandated: any team requesting new hires or budget must first prove AI cannot do the job. The logic has inverted from “why not hire” to “why not let AI do it first.”
Task-Switching Humans: Martin Fowler’s team categorizes human-AI collaboration into four modes:
- Human out of the loop: Low-risk, reversible tasks → Agent runs autonomously
- Human in the loop: High-risk, irreversible tasks → human review and control required
- Human on the loop: Mature workflows → human maintains rules and runs evaluations
- Agent flywheel: Self-improving autonomous execution
On-Demand Teams: Teams assemble around tasks, not permanent departments. Block restructured around “capability modules” (payment, lending, card issuance) combined by an intelligent orchestration layer based on customer scenarios. Moonshot AI calls this “division of labor, no boundaries” — functional boundaries do not define scope of work. Tezan restructured into a “pod + community” dual-track model: pods are cross-functional combat units delivering closed-loop outcomes; communities are horizontal skill pools (sales, product, code).
Elastic Contracts: When output and time decouple, fixed monthly salaries (pricing time) can become a penalty for high performers and a shelter for low performers. Tenex pays engineers by high-quality output volume (story points), not hours worked. Carta data: single-founder new companies in the US rose from 23.7% (2019) to 36.3% (H1 2025). The psychological contract is shifting from “loyalty in exchange for lifetime security” (relational) to “clear exchange, mutual freedom” (transactional).
The Governance Imperative: BCG/Columbia Business School research: 42% of executives admit AI adoption is “tearing the company apart”; 31% of employees admit actively sabotaging AI promotion. What tears organizations apart is not AI capability gaps — it is the gap between old hierarchical order being dismantled and new operating rules not yet being established. Liquid organizations do not reduce management difficulty; they push governance to a deeper level: clear task boundaries, unambiguous accountability, and explainable contribution distribution.
Trend 10: Role Reconstruction — From Supervisor to Architect
The most common question about AI and employment — “Will this job be replaced?” — is too coarse. Anthropic’s Economic Index provides granularity: mapping Claude usage to occupational tasks reveals that ~36% of occupations have at least one-quarter of their tasks showing AI involvement, but only ~4% have three-quarters or more of tasks covered by AI. AI does not consume entire jobs at once; it enters specific tasks within jobs.
The Task Recomposition Pattern:
A marketing operations role, decomposed:
- Tasks AI can handle today: Competitor information gathering, event data organization, first-draft report writing, basic data cleaning, standard customer service responses
- Tasks requiring human judgment: Reading user sentiment, coordinating channel relationships, calibrating brand tone, handling emergent risks
Usage Mode Shift (Anthropic Economic Index update):
| Mode | Initial Report | Sep 2025 Update |
|---|---|---|
| Augmentation (learning, verifying, iterating together) | 57% | Declining |
| Automation (model completes task directly) | 43% | Rising |
| Directive conversation share | 27% | 39% |
| Enterprise API automation mode | — | 77% |
Users increasingly trust AI with complete tasks: “You handle this first, I’ll verify.”
WEF Future of Jobs Report 2025 Projections (by 2030):
| Metric | Value |
|---|---|
| Jobs structurally impacted | ~22% of current roles |
| New jobs created | 170 million |
| Jobs displaced | 92 million |
| Net gain | 78 million |
| Core skills changing | ~39% |
From Task Doer to Agent Boss:
Microsoft’s Work Trend Index introduces the concept of the Frontier Firm and the provocative claim: “Every employee becomes an agent boss.” Key data:
- 81% of leaders expect Agents to be moderately or extensively integrated into enterprise AI strategy within 12–18 months
- 82% believe digital labor can expand employee capacity
- 83% believe AI will enable employees to handle complex work at earlier career stages
A researcher might manage one Agent scanning papers daily, one organizing data, and one writing briefings. An operations person might have one Agent doing competitor monitoring, one generating campaign materials, and one watching for data anomalies. You may not manage people, but you are managing a team of digital labor.
The Six Architectural Tasks (not six job titles — six task dimensions):
| Task Dimension | Core Question | Key Activities |
|---|---|---|
| Workflow Architecture | Where does AI plug into business processes? | Division of labor, sequencing, handoffs |
| Agent Orchestration | How do multiple Agents divide, collaborate, and hand off? | Multi-agent system design |
| Context Engineering | What knowledge, data, rules, and history does AI need? | Enterprise knowledge organization, customer data, project retrospectives |
| Evaluation & Trust | Is the output accurate, compliant, unbiased, traceable? | Quality assessment, compliance, IP protection |
| ROI Measurement | How much time, conversion, or rework was saved? | Business impact quantification |
| Learning Operations | How are templates, cases, and processes captured and reused? | Continuous improvement flywheel |
From Supervisor to Architect — Real-World Archetypes:
Architect capability is not abstract. It exists in roles across industries: film producers (deciding audience and tone), editors-in-chief (deciding story angles and fact-checking boundaries), general contractors (owning regulations, timelines, supply chains), chief physicians (orchestrating treatment plans and risks), game producers (balancing experience, retention, commercialization), and lab PIs (judging whether a research question is worth continued investment). These roles look different on the surface but share the same core competency: defining intent, setting constraints, and adjudicating evidence.
The Three Questions That Define Irreplaceability in the AI Era:
- Should we do this? (Intent definition)
- What does “done well” look like? (Constraint setting)
- What counts as evidence of success? (Evidence adjudication)
The ability to execute specific steps ranks last. The stronger AI becomes and the more efficient the execution layer, the larger the impact radius of architectural decisions.
The 2026 Summary: Execution capability is becoming cheap. Architectural capability is becoming expensive. Those who can define problems, organize tools, design systems, and bear consequences — they are the ones most likely to stand at the center of human-machine symbiotic organizations.
Cross-Cutting Themes and Implications
Theme 1: The Frontier Has Moved Outside the Model
Across all 10 trends, a consistent pattern emerges: the competitive frontier is shifting from what the model can do internally to what environment, rules, and systems surround the model. Harness Engineering (Trend 4) makes this explicit, but it echoes everywhere:
- Online Evolution (Trend 1): Context quality and infrastructure matter more than parameter count
- Multimodal Cognition (Trend 2): Memory, preferences, and safety boundaries are external systems
- Scientific Intelligence (Trend 3): Data access pathways and experimental infrastructure determine AI4S success
- Trusted Execution (Trend 6): Security and identity are protocol-level concerns, not model-level patches
- Intelligence as a Service (Trend 7): Value is in the packaging (workflows, outcomes, roles), not the raw tokens
- Liquid Organization (Trend 9): Structure is defined by rules and protocols, not org charts
- Role Reconstruction (Trend 10): Architecture is about designing systems, not executing within them
Theme 2: From Resources to Outcomes
The entire commercialization arc — from Token pricing to Agent role-based pricing (Trend 7), from attention economy to outcome economy (Trend 8), from employment contracts to contribution-based elastic contracts (Trend 9) — points in one direction: the unit of value is shifting from input consumption to output delivery.
Theme 3: The Identity and Trust Imperative
As Agents become autonomous actors (Trend 6, Trend 8), the absence of verifiable identity, traceable action chains, and enforceable permission boundaries becomes the single greatest systemic risk. The regulatory convergence across China, the EU, and global standards bodies in 2026 signals that trust infrastructure is the bottleneck — and the opportunity.
Theme 4: The Human Role Is Not Diminishing — It Is Concentrating
Every trend ultimately returns to the same question: what is left for humans to do? The answer, consistently, is direction-setting under uncertainty. AI excels at execution within defined parameters. Humans retain the comparative advantage in defining which parameters matter, what counts as success, and when to change course. The human role becomes more concentrated, more consequential, and more architectural — but it does not disappear.
Key Data at a Glance
| Metric | Value | Source |
|---|---|---|
| Frontier model context learning task completion | ~17% | CL-bench (Tencent) |
| ARC-AGI-3 performance (all frontier models) | <1% | ARC-AGI-3 benchmark |
| ARC-AGI-3 human performance | 100% | ARC-AGI-3 benchmark |
| AI production code at Anthropic | >80% | Anthropic disclosure |
| SWE-bench Verified improvement (Jun 2024 → Apr 2026) | 49% → 87.6% | SWE-bench |
| Claude Code revenue ramp | $0 → $2.5B (9 months) | Anthropic |
| AI-written code at Tencent (CodeBuddy users) | >50% | Tencent |
| YC W25 startups with >95% AI codebase | 25% | Y Combinator |
| Token calls priced <$4/M | ~76% of total volume | OpenRouter |
| Enterprises deploying AI Agents in core processes | >45% | IDC |
| Average ROI of enterprise AI Agent deployment | 171% | IDC |
| AI-discovered drugs in clinical development | ~170 | Industry estimates |
| AI-generated fake references (per 10K papers) | 4.0 (2023) → 56.9 (2026) | The Lancet |
| Cursor team size at $300M ARR | ~60 people | Cursor |
| Traditional team size for $300M ARR | 500–1,000 people | Industry norms |
| US single-founder new companies (2019 vs H1 2025) | 23.7% → 36.3% | Carta |
| Anthropic annualized revenue / growth team size | $14B / ~40 people | Anthropic |
| Global net job change projected by 2030 | +78 million | WEF Future of Jobs 2025 |
| Executives reporting AI “tearing company apart” | 42% | BCG / Columbia Business School |
| XtalPi monthly data production | 50K+ reaction yields, 300K process data | XtalPi |
| Isomorphic Labs cumulative external funding | ~$2.7B | Isomorphic Labs |
| Lila Sciences total funding | $550M | Lila Sciences |
