AI news

The Rise of AI Agents: A New Turning Point for AI+Software Development?

Honghao Wang

— 3 min read

LLM Native Development Era

Large language models (LLMs) are transforming R&D workflows — from assisted programming to autonomous coding agents — shifting AI from a supporting tool to a driver of productivity. But have we truly entered a “native development era,” or are we still climbing the slope?

Recently, InfoQ’s Geek Talks, in collaboration with AICon, hosted a live session featuring:

  • Wu Chaoxiong — Senior Product Manager, Ping An Technology
  • Yan Zhijie — Senior Architect, Baidu
  • Du Pei — Client Architect, Autohome

They discussed new paradigms in software development at the eve of AICon Global Artificial Intelligence Development & Application Conference 2025 Beijing (December).

📅 Full conference schedule:

https://aicon.infoq.cn/202512/beijing/schedule

---

Key Takeaways from the Live Discussion

1. Current State of AI in Development

  • Testing: Still auxiliary — boosts efficiency but far from replacing humans end-to-end.
  • Prompt Engineering: Fundamentally role-playing — setting domain-specific roles and constraints to guide model behavior.
  • Coding Agents: Early steps toward general-purpose intelligent agents, capable of multistage reasoning and autonomous execution.
  • Maintainable Code: The real goal is production-ready, maintainable software — not just “code that runs.”
  • Tool Choice vs. Outcomes: Using AI does not automatically mean better results; not using AI doesn’t imply worse ones.

🎥 Replay of the session:

https://www.infoq.cn/video/KGWqbzH6IKGhgoUiKDEi

---

Perspectives from the Panel

Wu Chaoxiong: Public vs. Developer Perception

Many think “AI coding” is just advanced autocomplete — not a paradigm shift.

In reality, AI can excel at isolated, well-structured tasks (the “fire”) but struggles with complex legacy systems (the “seawater”).

For non-programmers, LLMs can enable tasks previously impossible. For career coders, change is underway but incomplete.

---

  • Beyond IDEs: Examples include Devin, SWE Agent, and CLI tools like Claude Code integrated in DevOps pipelines.
  • Teams reporting 50%+ AI-generated code show AI’s deep workflow penetration.
  • For non-developers, this is already a paradigm shift.
  • For professionals, we’re nearing a turning point, but ultimate transformation requires stable performance and deeper AI capabilities.

---

Du Pei: Varied Depth of AI Use

  • Shallow usage: Simple Q&A, code snippets.
  • Deep usage: Process-oriented intelligent agents, dynamic cross-platform migration (e.g., Android → iOS).
  • AI excels at standardized tasks with clear rules, but struggles in complex scenarios (e.g., 3D modeling).

---

Representative AI Application Scenarios

1. Design → Code Conversion

  • Early attempts (2023) had high hallucination rates.
  • Multimodal models improved results by parsing image + design context (e.g., with MasterGo).
  • Achieved 80–90% usability — still requires manual review for pixel-perfect precision.

2. Cross-Platform Code Migration

  • Migrating logic from H5, mini-programs, or frameworks achieves 70%+ quality, but requires manual compliance checks.

3. Code Review & Bug Detection

  • AI detects issues using combined sets of established project rules.
  • Reduced testing-stage bugs by 30–40%.

---

Capability Growth — Roles in Transition

AI-era Competencies

  • Prompt Engineering — Structured, role-defined instructions.
  • Knowledge Engineering — Feeding AI with codified business processes and standards.
  • System Thinking — Understanding constraints, architecture, and exception handling.
  • Collaboration Skills — Communicating tasks effectively to AI.
  • Taste / Product Vision — Higher-level judgment that AI cannot replicate.

---

Amplified Roles

  • QA Engineers: Shift focus from how to test → what to test.
  • Product Managers: Move beyond logic articulation → specify functional & non-functional requirements with architectural awareness.
  • Full-Stack Engineers: Use AI to cross boundaries — e.g., front-end devs leveraging AI for backend SQL queries.
  • Architects: Orchestrate system-level AI integration — AI covers small tasks; humans oversee overall coherence.

---

Intelligent Agents vs. AI Assistants

AI Assistants:

  • Support single tasks (autocomplete, Q&A).

Intelligent Agents:

  • Autonomously execute multistage workflows (coding → testing → deployment).
  • Operate with closed-loop capabilities — observe, reason, act.
  • Examples: Devin, SWE Agent.

---

Deployment Challenges — First Wall AI Hits

  • Stability: Variance in output quality obstructs habit change.
  • Trust Building: Users remember failures more than successes.
  • Prompt Quality: Vague requests lead to poor results.
  • Human Oversight Requirement: Complex orchestration logic still needs experts.

---

Practical Adoption Strategy

  • Start small — component-level AI integration.
  • Define standards & templates — reduce prompt ambiguity.
  • Gradually expand scope — build trust from proven successes.
  • Use plugin-based approach — embed AI in existing workflows before attempting unified platforms.

---

Linking R&D & Monetization Platforms

Platforms like AiToEarn官网 illustrate an open-source model integrating:

  • AI generation
  • Cross-platform publishing (Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X)
  • Analytics
  • Model ranking (AI模型排名)

While designed for content creators, such orchestration principles inspire enterprise AI workflows — connecting ideas, automation, and measurable value.

---

Executive Summary — Strategies for Professionals in the AI Era

For Technologists & Managers:

  • Master Prompt + Knowledge Engineering to ensure AI outputs align with real business needs.
  • Identify standardized, decomposable tasks where AI adds maximum value.
  • Leverage multimodal models for richer, more accurate context interpretation.
  • Explore plugin-based ecosystems for gradual, low-risk AI integration.
  • Focus your role on decision-making & orchestration — areas AI cannot replace.

For Organizations:

  • Drive AI adoption via culture and guidance, not enforced quotas.
  • Measure success by output quality, not AI usage percentage.
  • Build trust through stability, quick iterations, and transparency in AI-driven workflows.
  • Prepare for future cross-modal capabilities — text, image, video, and behavior recognition.

---

📎 References & Links

---

If you’d like, I can now produce a condensed, bullet-point “Action Playbook” for thriving in AI-assisted development — focused purely on actionable career steps. Would you like me to proceed?

Read more

Translate the following blog post title into English, concise and natural. Return plain text only without quotes. 哈佛大学 R 编程课程介绍

Harvard CS50: Introduction to Programming with R Harvard University offers exceptional beginner-friendly computer science courses. We’re excited to announce the release of Harvard CS50’s Introduction to Programming in R, a powerful language widely used for statistical computing, data science, and graphics. This course was developed by Carter Zenke.