Tencent Research Institute releases "AI Coding Observation Report 2.0"

After the Abundance:AI The "Coding Observation Report 2.0" is a report launched by Tencent Research Institute.AIThis in-depth research report on lenses, authored by Cao Shiyi, Yu Yi, and Yuan Xiaohui, covers the period from the second half of 2025 to the first quarter of 2026. The report is based on the first edition of [the report's title/article/etc.].AI Nine months after the release of the "Coding Non-Consensus Report," the system verified the implementation of seven non-consensus principles and extracted six structural insights. The core proposition is: When...AI After Coding overcomes the two capability thresholds of "assisted completion" and "autonomous collaboration", code generation has entered an era of abundance, and the scarcity has shifted from "how to write" to "how to verify, how to define specifications, and how to operate continuously".

• Studying veinsThe first version of the report, dated July 2025, proposed...AI Coding is a general term.AgentThe second edition addressed seven points of contention, including the "prior battlefield." Nine months later, most judgments had become consensus, but new structural problems emerged. The second edition no longer debated these points.AIThe question of "Can we program?" is a question that seeks to answer the question, "After abundance comes scarcity."
• Time spanFocusing on the second half of 2025 to the first quarter of 2026, covering...Claude Key events included the release of Opus 4.5 (the second hurdle), the reversal of METR experiment results, the surge in Cursor's valuation, the explosion of the Skills ecosystem, and three "Anthropic Days".
• Research MethodsBased on laboratory technical reports, third-party benchmarks (SWE-bench/Vellum), industry research (a16z/Pragmatic Engineer/GitClear), and enterprise practices (Anthropic/Open).AI/Cursor) and security incident analysis, using "AIThe lens penetrates noise and extracts verifiable structural trends.

• Models are converging rapidly, but the gap at the forefront remains.The six major business models were compressed to a 1% range on SWE-bench Verified.open sourceWith Qwen3-Coder reaching the 80% mark, "which model to choose" is no longer a core decision for most companies. However, Anthropic released Mythos Preview (93.9%, not publicly available) and Opus 4.7 (87.6%, publicly available) within 9 days in April 2026. The dual-track mechanism of "continuous breakthroughs in internal capabilities + phased differentiated de-weighting of public releases" indicates that a new gap is widening between the capability reserves of cutting-edge laboratories and publicly available models.
• AgentThe original direction of tool evolution has become the convergence direction.In terms of form, Cursor 3, Codex App, and Google Antigravity are transforming IDEs from "code editors +..."AIThe plugin has been upgraded to...Agent"Orchestration platform + code view". At the interface level, CLI wins.AgentThe inner loop (gh command approximately 200 tokens vs GitHub MCP server 55,000 tokens, a difference of 275 times), MCP retreats to the enterprise outer loop, and Skills uses SOPs to encapsulate and become the preferred interface for non-developers.AgentThe best working environment is an orchestration platform, and the best capability interface is code and SOPs.
• Scaling up code generation presents a new bottleneck for verification.The SWE-bench score of 87.6% indicates a core bottleneck in "how to achieve" exiting the bottleneck. New bottlenecks appear at both ends: forward lies specification definition (KTH experiments show...).AI(A complete bootstrap code with a 926-word English specification is available), followed by verification and maintenance (Veracode found 45%).AIThe code task contained known vulnerabilities; GitClear analysis of 211 million lines of code showed a 30-41% increase in technical debt. The next wave of value creation will not be in better code generation, but in better specifications, verification, and maintenance infrastructure.
• Product creation has zero barriers to entry, but taste and operational expertise are becoming increasingly rare.YC W2025 batch: 25% of startups, 95%+ code provided.AIThe proportion of solo founders increased from 23.7% to 36.3%, and Base44 was acquired by Wix for $80 million within six months. "Prototype walls" are prevalent.AIfastAfter the MVP is generated, security, scalability, and edge cases emerge, and maintenance costs rise from $20/month to $200/month.AIEliminating the "from zero to prototype" barrier has made distribution, operation and maintenance, compliance, and taste the new scarce resources.
• SaaS is not dead, it is being redistributed.Three "Anthropic Days" targeted middleware SaaS: Cowork's release caused FactSet to drop 10%, and the COBOL blog caused IBM to drop 13.2% (the largest single-day drop in 25 years).Claude Design caused Figma to fall 6.89%. The victims were all single-function middleware layers that "wrapped APIs as paid UIs." During the same period, Cursor's valuation jumped from $29.3 billion to $50 billion, its Skills catalog exceeded 2,600, and both its platform and self-built layers grew significantly. The pricing unit shifted from "per seat" to "per output."
• What to do and who can do it: Developers are being redefined in both directions.The roles are changing: developers are shifting from "writers" to "programmers," and staff+ engineers (63.5%) are the most heavily involved.AgentFor users, judgment and system understanding have become core skills. The landscape is also changing: non-developers are entering the field as "builders" for the first time, with over 50% of Epic Games members doing so.Claude The code is provided by non-developers, and Skills allows non-developers to write their first Markdown code within 10 minutes.AgentSkills. Employment flows through three levels: entry-level coding is reduced, mid-level technical project managers are added (LinkedIn +250%), and high-level architecture and product definition are amplified.

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

• The first hurdle (mid-2024)Claude 3.5 Sonnet):AIIt has evolved from an auxiliary code completion tool to a reliable code generation assistant, expanding its capabilities from Copilot-style inline completion to Cursor-style full-stack writing.
• The second hurdle (end of 2025 - first half of 2026, Opus 4.5/4.7):fromSimpleFrom task assistance to stable autonomous collaboration, Opus 4.5 achieved a breakthrough of 80.9% in SWE-bench Verified for the first time (reaching 80.9%), with token consumption decreasing by approximately 65% compared to the previous generation.
• Five-dimensional evidence convergenceTechnology (SWE-bench breakthrough), Products (Plan Mode/Multiple)Agent/1M context), user experience ("Claude The "Christmas" phenomenon), opinion leaders (Karpathy went from "the model is garbage" to "Opus 4.5 is 10 times stronger"), business (Claude Code ARR grew from zero to $2.5 billion, and Anthropic's valuation reached $380 billion.
• METR reverseThe first RCT in early 2025 showedAIThe experiment slowed developers down by 19%, but a follow-up experiment in February 2026 reversed this to 18% faster (CI -38% to +9%), with 30-50% of developers rejecting the "no" approach.AI"condition.

Chapter 2 Models and Driving Engineering: Convergence, Differentiation, and Co-evolution

• Commercial convergenceSix leading models (Opus 4.5/4.6,Gemini 3.1 Pro, MiniMax M2.5, Kimi K2.6,GPT-5.4) Compressed to a 1 percentage point range on SWE-bench Verified.
• Frontier DifferentiationOpus 4.7 jumped 6.8 percentage points to 87.6%, breaking the five-month convergence range; Mythos Preview reached 93.9% but was not disclosed, forming a dual-track mechanism.
• Harness EngineeringWhen models converge, competitiveness shifts to mastering the framework. On SWE-bench Pro, changes in scaffolds cause 22 times the score fluctuation as model replacements. The skill evolution path...Prompt wordsFrom engineering (2022-24) to context engineering (2025) and then to driving engineering (2026).
• manyAgentArrangementFrom Sub-agents (a brain directing the hands and feet) toAgent Teams (each team has its own responsibilities), Kimi K2.6 pushes up to 300 sub-agents running in parallel for 12 hours continuously.
• Source code leak inspirationMarch 2026Claude The code unexpectedly exposed 512,000 lines of TypeScript source code, revealing over 40 discrete capabilities, including the Tool System and K.AIDesigns such as ROS daemon process and autoDream memory integration.

Chapter 3: Reshaping the Tool EcosystemAgent-First, CLI and Skills

• Agent-First TransformationThe IDE's positioning has shifted from "code editor +..."AIThe plugin has been upgraded to...Agent"Orchestration platform + code view". Cursor 3.0 is released.Agents Window supports parallel execution across repositoriesAgentOpenAI The Codex App reached 4 million weekly active users within two and a half months of its launch.
• CLI vs MCPCLIAgentSince it's a native language (training data contains massive amounts of shell commands), MCP requires an additional layer of abstraction. Industry convergence to hierarchical routing: CLI wins local.fastIteration (inner loop), MCP wins cross-system coordination within enterprises (outer loop).
• Skills Ecosystem:Agent Skills are organized into folders (SKILL.md + optional scripts).AgentThe "Onboarding Handbook" employs a progressive disclosure and on-demand loading approach. It utilizes a three-tier architecture: Skills (SOP layer, when/how to do it) encapsulates MCP (Integration layer, connecting to external systems) and CLI (Execution layer, infrastructure operations). Skills allow non-developers to directly become employees from the outset.AgentThe author's ecosystem doubled to over 2600 curated entries in one quarter.

Chapter 4 When Builders Are No Longer Scarce: Bottleneck Migration, Prototype Walls, and Track Dissolution

• Bottleneck migrationIn the past, the bottleneck was in "writing and implementing".AIAfter elimination, the bottleneck shifts to the "defining specifications" and "verifying and maintaining" ends. Karpathy proposes Software 1.0 (Code) → 2.0 (Weights) → 3.0 (...).Prompts/Specs) paradigm evolution.
• Prototype WallThe first week was exciting (AIfastMVP generation → Week 3 concerns (security, scalability issues emerge) → Month 2 abandonment (maintenance costs exceed expectations). Addy Osmani raises the "70% problem":AIThe code may appear to be 70% correct, but completing the remaining 30% could cost more than writing it from scratch.
• Track melting:AI Coding as a distinct category is beginning to blur its boundaries; programming capabilities (reading and writing files + executing commands + iterative repair) are inherently equivalent to general-purpose programming.AgentCapability stack.Claude Code→Agent The SDK→Cowork→Channels path is clear, and multiple entry points (IDE/Terminal/IM/Web/Desktop) are converged.

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

• SaaS reallocationThe middleware layer, the "complexity encapsulation layer" SaaS, is being impacted, while the platform layer (Cursor/Anthropic) and the minimalist self-built layer (Skills ecosystem) are growing stronger. IDC predicts that pure seat billing will be obsolete by 2028, and the pricing unit will migrate from "per seat" to "per outcome/consumption".
• Three new attack surfacesLiteLLMSupply chain attack (AIThe tool itself becomes a target of attack), the Axios npm supply chain incident (old paradigm +AIAcceleration), Vercel/Context.ai identity hijacking (AIThe OAuth tool was compromised, and the attacker was...AISignificantly accelerated).
• Offensive and defensive symmetry declineNicholas Carlini usedClaude Code discovered a 23-year-old vulnerability in the Linux kernel; Mythos Preview discovered thousands of zero-day vulnerabilities (the cost of discovering the 27-year-old TCP SACK vulnerability in OpenBSD was less than $50). The cutting-edge lab adopts a new paradigm of "dual-track demotion + authentication access".

Chapter Six: Looking Towards the Future: Role Transformation, Non-Developer Entry, and Job Mobility

• Role TransformationDevelopers are shifting from "writers" to "organizers," with time allocation moving upstream (defining specifications, designing constraints, evaluating deliverables, and managing collaboration). The Amjad Masad three-quadrant approach: downwards (system programming/embedded systems/safety-critical), in the middle (full-stack/CRUD, the most dangerous), and upwards (product design/business strategy).
• Education system responseStanford offers "The Modern Software Developer" course, encouraging students not to write code; MIT offers "No Code and..." Agentic AI"Specialized courses; 62% of US university CS enrollments declined, but..."AIA surge of expertise.
• Non-developers allowed to participateEpic Games accounts for over 50%Claude Code is used by non-developers; Block's non-engineering employees built their own MCP server; Bolt.new's 60-70% of users are non-traditional developers.
• One-person companies becoming mainstreamYC's solo founder ratio has risen to 36.3%, and the idea of "10 people doing the work of 100 people" has gone from a prediction to the norm. Cursor achieved 100 million ARR with 20 people, and Bolt.new achieved 20 million ARR with 15 people in 2 months.
• Three-tiered employment mobilityThe value of senior management (structure/judgment) is increasing, while the value of middle management (management) is rising.AgentThe technical project manager has been added, and the underlying (primary code) has been compressed.

Tencent Research Institute's "After Abundance:AI The Coding Observation Report 2.0 points out that, with...Claude Models like Opus 4.7 cross the threshold of autonomous collaboration.AI Coding has entered an era of code abundance, with scarcity shifting from "how to write" to "how to define specifications, verify and maintain, and continuously operate." Developers have transitioned from writers to orchestrators, and non-developers are entering the field as builders for the first time through Skills, with tools becoming increasingly comprehensive.Agent-First-based SaaS middleware is redistributed into two poles: a platform layer and a minimalist self-built layer.AIAt the same time, lowering the threshold for attack and defense poses new challenges to supply chain security.

Original link:After the bounty:AI Coding Observation Report 2.0 |AI Lens Series Research