Agents Build Software

1. Agents Build Software 演讲人：张海⻰ Gru.ai 创始人 Coding.net 创始人 OsChina 联合创始人

2.

3. The Future Software Development Workflow

4. The Future of Software Development Product Design Arch Design Coding Debugging Testing Creative Works Routine Works Handled by Humans Mostly Handled by Coding Agents Operating

5. Two Types of Coding Agents For Citizen For Enterprise

6. The Future of Software Development Can Lovable Maintain Sustainable Growth?

7. The Challenges of Coding Agents The Ultimate AI Developer Issue Reproduce Issue Prepare Patch Verify Changes PR

8. Bug Fix Agent - Keeping on SOTA Rank No.1 on SWE-bench verified leaderboard twice on 2024.08 and 2024.12. SWE-bench is the most recognized benchmark for coding agents. https://www.swebench.com/ Bug fix agent is still premature, not ready for commercialization.

9. The Challenges of Coding Agents The Ultimate AI Developer Issue Reproduce Issue Prepare Patch Verify Changes Essential Capabilities to Complete the Process Requirement Understanding Read/Edit Files Terminal Use Code Analysis Setup Env Browser Use PR

10. Current Solutions Human High Touch Human Low Touch Human Assisted Missing Capabilities Focus on Vertical Scenarios Requirement Understanding Read/Edit Files Terminal Use Requirement Understanding Read/Edit Files Terminal Use Code Analysis Setup Env Browser Use Code Analysis Setup Env Browser Use

11. Current Solutions

12. Coding Agents in Your Workflow

13. Enterprise Coding Agent Challenges AI currently has poor ability to understand the customer’s engineering environment. Low execution accuracy leads to a limited range of applicability.

14. Unit Test Agent - Current Issues 60% unit test coverage significantly reduces bugs, but adds 30% more developer time.

15. Unit Test Agent - Current Issues Two Functions of Unit Testing: 1. Improve the quality of code writing 2. Regression testing

16. Unit Test Agent Scenario 1: Agent auto-completes unit tests on new code New Code Submitted Agent Completes Unit Test Code Review

17. Unit Test Agent Scenario 2: Agent submits batch unit tests on existing code Agent Analyze Existing Code Agent Agent Submit Submit Unit Unit Tests Tests in in in Batch Batch Agent Agent Submit Submit Unit Unit Tests Tests in Batch Batch Code Review

18. Unit Test Agent Scenario 3: If any obvious errors are detected, you will be notified.

19. E2E Test Agent - Current Issue Too Much Manual Work

20. E2E Test Agent - Current Issue Hard to Maintain Testing Code

21.

22. E2E Test Agent - Challenges

23. More Coding Agents Refactor Bug Fix Code Review

24. Agents Contribute to Your Repos

25. Engineering Approaches to Building a Coding Agent

26. How do We Build Agents to Solve Real life Problems Problem Definition Evaluation Work with LLMs Curated Context Agent OS

27. Basic Characteristics of LLM

28. Agent Evaluation

29. Work with LLMs Model Agnostic Mult-Models FineTuned Models

30. Curated Context Adjust for Languages and Frameworks Build into the Workflow

31. Agent OS

32. Agent OS - Debug Console

33. Agent OS - Open Source gbox is an open source project that provides a self-hostable sandbox for Agents. https://gbox.cloud Terminal Browser MCP Desktop iOS/ Android File

34.

35. Embrace Agents in Workflow 探索 AI 应用边界 Explore the limits of AI applications