Tencent Research Institute releases "AI Coding Observation Report 2.0"

"After Abundance: AI Coding Observation Report 2.0" is a deep research report in the AI Lens series released by Tencent Research Institute, written by Cao Shiyi, Yu Yi, and Yuan Xiaohui, covering the second half of 2025 to the first quarter of 2026. The report stands on the first edition"AI Coding Non-Consensus Report"nine months after its release, systematically verifies the implementation status of 7 non-consensus points and distills 6 structural insights. The core proposition is: once AI Coding crosses the two capability thresholds of "assisted completion" and "autonomous collaboration," code generation enters an era of abundance, and scarcity shifts from "how to write" to "how to verify, how to define specifications, how to continuously operate."

Report Background and Positioning

• Research Context: The first edition of the report in July 2025 proposed 7 non-consensus points, including "AI Coding is the a priori battlefield for general Agents." Nine months later, most judgments have become consensus, and new structural problems have emerged. The second edition no longer debates "whether AI can program," but asks "after abundance, what becomes scarce."
• Time Span: Focuses on 2025H2 to 2026Q1, covering key events such as the release of Claude Opus 4.5 (the second threshold), the reversal of METR experimental conclusions, Cursor's valuation surge, the explosion of the Skills ecosystem, and three "Anthropic Day" events.
• Research Method: Based on laboratory technical reports, third-party benchmarks (SWE-bench/Vellum), industry research (a16z/Pragmatic Engineer/GitClear), enterprise practices (Anthropic/OpenAI/Cursor), and security incident analysis, it uses the "AI Lens" to penetrate noise and extract verifiable structural trends.

6 Structural Insights

• Models accelerate toward convergence, but the frontier gap does not narrow: The six major commercial models have compressed into a 1-percentage-point range on SWE-bench Verified, and open-source Qwen3-Coder has caught up to the 80% tier; "which model to choose" is no longer a core decision for most enterprises. But Anthropic released Mythos Preview (93.9%, not public) and Opus 4.7 (87.6%, public) within 9 days in April 2026. The dual-track mechanism of "continuous internal capability breakthroughs + phased differentiated de-weighting of public releases" indicates that a new gap is opening between the capability reserves of frontier laboratories and public models.
• Agent-native becomes the converging direction of tool evolution: At the form level, Cursor 3, Codex App, and Google Antigravity upgrade the IDE from a "code editor + AI plugin" to an "Agent orchestration platform + code view." At the interface level, CLI wins the Agent inner loop (the gh command is about 200 tokens vs the GitHub MCP server's 55,000 tokens, a 275-fold difference), MCP retreats to the enterprise outer loop, and Skills, packaged with SOPs, become the preferred interface for non-developers. The best working environment for an Agent is an orchestration platform, and the best capability interface is code and SOPs.
• Code generation scales up, verification becomes the new bottleneck: SWE-bench 87.6% means "how to implement" has exited the core bottleneck. The new bottleneck appears at both ends: forward is specification definition (the KTH experiment shows AI can fully bootstrap code from a 926-word English specification), and backward is verification and maintenance (Veracode found that 45% of AI code tasks contain known vulnerabilities, and GitClear, after analyzing 211 million lines of code, found technical debt increased by 30-41%). The next wave of value creation is not in better code generation, but in better specifications, verification, and maintenance infrastructure.
• Product building has zero barrier to entry, while taste and operations gradually become scarce: YC W2025 batch: 25% of startups have 95%+ of code generated by AI, the proportion of solo founders rose from 23.7% to 36.3%, and Base44, built by a single person in 6 months, was acquired by Wix for $80 million. The 'prototype wall' is widespread: after AI rapidly generates an MVP, security, scalability, and edge cases emerge, and maintenance costs rise from $20/month to $200/month. AI eliminates the 'from zero to prototype' threshold, but distribution, operations, compliance, and taste become the new scarcities.
• SaaS has not died; it is being redistributed: Three targeted 'Anthropic Day' strikes hit middle-layer SaaS: the Cowork launch caused FactSet to fall 10%, the COBOL blog caused IBM to fall 13.2% (its largest single-day drop in 25 years), and Claude Design caused Figma to fall 6.89%. The victims are all single-function middle layers that 'wrap APIs in a paid UI.' During the same period, Cursor's valuation jumped from $29.3 billion to $50 billion, the Skills directory exceeded 2,600 entries, and both the platform layer and the build-your-own layer grew stronger at the two poles. The unit of pricing is shifting from 'per seat' to 'per output.'
• What to do and who can do it: developers are being redefined in both directions: 'What to do' is changing: developers are shifting from 'writers' to 'orchestrators,' Staff+ engineers are 63.5% the heaviest Agent users, and judgment and systems understanding have become core skills. 'Who can do it' is also changing: non-developers are entering for the first time as 'builders,' over 50% of Claude Code usage at Epic Games comes from non-developers, and Skills let non-developers write their first Agent capability in markdown within 10 minutes. Employment is flowing across three layers: entry-level coding at the bottom is being compressed, technical project manager roles are being added in the middle (LinkedIn +250%), and architecture and product definition at the top are being leveraged and amplified.

Detailed Report Contents

Chapter 1 The Moment of Qualitative Change: Two Capability Thresholds and Five-Dimensional Evidence

• The first threshold (mid-2024, Claude 3.5 Sonnet): AI jumped from assisted completion to a reliable code generation assistant, and its capability boundary expanded from Copilot-style inline completion to Cursor-style full-stack writing.
• The second threshold (late 2025-1H 2026, Opus 4.5/4.7): It crossed from assistance on simple tasks to stable autonomous collaboration. Opus 4.5 broke 80% on SWE-bench Verified for the first time (reaching 80.9%), and token consumption fell by about 65% compared with the previous generation.
• Five-dimensional evidence converges: technology (SWE-bench breakthrough), product (Plan Mode/multi-Agent/1M context), user experience (the 'Claude Christmas' phenomenon), opinion leaders (Karpathy shifted from 'models are just garbage' to 'Opus 4.5 is 10 times stronger'), and business (Claude Code ARR went from zero to $2.5 billion, and Anthropic's valuation reached $380 billion).
• METR reversal: In early 2025, the first RCT showed AI made developers 19% slower; in February 2026, a follow-up experiment reversed this to 18% faster (CI from -38% to +9%), and 30-50% of developers refused the 'no AI' condition.

Chapter 2 Models and Harness Engineering: Convergence, Divergence, and Coevolution

• Commercial convergence: Six leading models (Opus 4.5/4.6, Gemini 3.1 Pro, MiniMax M2.5, Kimi K2.6, GPT-5.4) compressed into a 1-percentage-point range on SWE-bench Verified.
• Frontier divergence: Opus 4.7 jumped 6.8pp in one leap to 87.6%, breaking the 5-month convergence range; Mythos Preview reached 93.9% but is not public, forming a dual-track mechanism.
• Harness Engineering: When models converge, competitiveness shifts to the harness framework. On SWE-bench Pro, score fluctuations caused by scaffold changes are 22 times those caused by model replacement. The skill evolution path goes from prompt engineering (2022-24) to context engineering (2025) and then to harness engineering (2026).
• Multi-Agent orchestration: From Sub-agents (one brain directing hands and feet) to Agent Teams (a team each performing its own duties), Kimi K2.6 pushes to 300 sub-agents running in parallel and 12 hours of continuous execution.
• Source code leak inspiration: In March 2026, Claude Code accidentally exposed 512,000 lines of TypeScript source code, revealing designs such as a Tool System with 40+ discrete capabilities, the KAIROS daemon, and autoDream memory consolidation.

Chapter 3 The Reshaping of the Tool Ecosystem: Agent-First, CLI, and Skills

• Agent-First transformation: The IDE positioning upgrades from "code editor + AI plugin" to "Agent orchestration platform + code view." Cursor 3.0 launches the Agents Window to support parallel operation of multiple Agents across repositories; OpenAI Codex App reaches 4 million weekly active users two and a half months after launch.
• CLI vs MCP: CLI is a native language for Agents (training data contains massive shell commands), while MCP requires an additional abstraction layer. The industry converges on layered routing: CLI wins local rapid iteration (inner loop), MCP wins enterprise cross-system coordination (outer loop).
• Skills ecosystem: Agent Skills uses a folder (SKILL.md + optional scripts) as an Agent's "onboarding manual," adopting progressive disclosure with on-demand loading. Three-layer architecture: Skills (SOP layer, when to do/how to do) encapsulates MCP (integration layer, connecting external systems) and CLI (execution layer, infrastructure operations). Skills enable non-developers to directly become Agent authors for the first time, and the ecosystem doubles in a quarter to 2,600+ curated directories.

Chapter 4 When Building Is No Longer Scarce: Bottleneck Migration, the Prototype Wall, and Track Dissolution

• Bottleneck migration: In the past, the bottleneck was "writing the implementation"; after AI eliminates it, the bottleneck migrates to both ends: "defining specifications" and "verification and maintenance." Karpathy proposes the paradigm evolution of Software 1.0 (Code) → 2.0 (Weights) → 3.0 (Prompts/Specs).
• Prototype Wall: First week excitement (AI rapidly generates an MVP) → third week concern (security and scalability emerge) → second month abandonment (maintenance costs exceed expectations). Addy Osmani proposes the "70% problem": AI code appears 70% correct, but completing the remaining 30% may cost more than writing it by hand from scratch.
• Track dissolution: The boundary of AI Coding as an independent category begins to dissolve; programming capability (reading and writing files + executing commands + iterative repair) is naturally equivalent to the general Agent capability stack. The path Claude Code → Agent SDK → Cowork → Channels is clear, with multiple entry points (IDE/terminal/IM/Web/desktop) converging.

Chapter 5 Landscape and Security: SaaS Reallocation and Three New Attack Surfaces

• SaaS reallocation: The middle layer, the "complexity encapsulation layer" SaaS, is under impact, while the platform layer (Cursor/Anthropic) and the minimalist self-built layer (Skills ecosystem) grow stronger at both poles. IDC predicts that pure seat-based billing will become obsolete before 2028, and pricing units will migrate from "per seat" to "per outcome/consumption."
• Three new attack surfaces: LiteLLM supply chain attack (AI tools themselves become attack targets), Axios npm supply chain incident (old paradigm + AI acceleration), Vercel/Context.ai identity hijacking (AI tool OAuth is compromised, and attackers are significantly accelerated by AI).
• Symmetric decline in attack and defense: Nicholas Carlini used Claude Code to discover a 23-year-old vulnerability in the Linux kernel; Mythos Preview discovered thousands of zero-day vulnerabilities (the cost of discovering OpenBSD's 27-year-old TCP SACK vulnerability was less than 50 dollars). Frontier labs adopt a new paradigm of "dual-track privilege reduction + identity verification admission."

Chapter 6 Facing the Future: Role Transformation, Non-Developer Entry, and Employment Mobility

• Role transformation: Developers are shifting from "writers" to "orchestrators," with time allocation migrating upstream (defining specifications, designing constraints, evaluating outputs, managing collaboration). Amjad Masad's three quadrants: going down (systems programming/embedded/safety-critical), staying in the middle (full-stack/CRUD, the most dangerous), going up (product design/business strategy).
• Response from the education system: Stanford offers the "The Modern Software Developer" course, encouraging students not to write code; MIT offers the "No Code and Agentic AI" specialized course; CS enrollment at 62% of U.S. universities is declining, but AI majors are surging.
• Non-developers entering the field: More than 50% of Claude Code usage at Epic Games comes from non-developers; non-engineer employees at Block build their own MCP servers; 60-70% of Bolt.new users are non-traditional developers.
• One-person companies becoming mainstream: The proportion of YC solo founders has risen to 36.3%, and "10 people doing the work of 100" has gone from a prediction to the norm. Cursor reached $100 million ARR with 20 people, and Bolt.new reached $20 million ARR with 15 people in 2 months.
• Three tiers of employment mobility: The value of the upper tier (architecture/judgment) is rising, a new middle tier (technical project managers managing Agents) has emerged, and the lower tier (junior coding) is being compressed.