← back
How Coding Agents change Software Development Forever - Hailong Zhang
Takeaway
Effective coding agents are narrow, evaluation-grounded, and built on a shared agent-OS abstraction — not monolithic 'software engineer' bots.
Summary
- Two coding-agent paradigms: synchronous (Copilot/Cursor inside IDE, mature since 2023) and asynchronous (GitHub-bot style, new in 2024) — both needed.
- Demo of 'Guru' agent that auto-generates unit tests triggered by PRs — over 50% of Guru's PRs get human-merged in production, handles 80% of unit tests in its own repo, top contributor by commit count.
- Building agents requires: a concrete bounded problem (unit-test, not 'software engineering'), eval dataset + harness, LLM selection per stage (Guru uses multiple frontier models, fine-tuned GPT-4 for test code), context engineering per language/framework.
- Built an internal 'agent OS' to share runtime/tools/context across many task-specific agents (refactor, E2E test, etc.).
coding-agentsunit-testsagent-os
Original description
In this talk, we explore the transformative role of coding agents in modern software development. We'll begin by examining the future software development workflow, highlighting how coding agents streamline processes and enhance productivity. Next, we'll discuss the process of working with an unit test agent (Test Gru) as an example. We'll also share some development experiences with Gru.ai agents to help you better understand how to leverage them to enhance efficiency. Join us to discover how coding agents are shaping the future of development.