Advances in Code Large Models: AIGC Impact on Software Development and Future Multi‑Agent Tools

Code large models have become mature technology products widely applied across industries; this presentation shares the latest research results of the Tongyi Lingma project.

The content is organized into four parts: (1) the fundamental impact of AIGC on software R&D, (2) opportunities and challenges of intelligent software development, (3) new human‑machine collaboration modes enabled by AI, and (4) changes in knowledge transmission during development.

AIGC's Fundamental Impact on Software R&D

Enterprise R&D efficiency is constrained by three core factors—personnel skill, collaboration overhead, and cost control. AIGC can improve skill gaps, reduce coordination friction by creating “super‑individuals,” and help lower costs, ultimately offering unprecedented opportunities for software engineering.

Intelligent Opportunities and Challenges

Four aspects are highlighted: individual efficiency gains through Copilot‑style interaction, collaborative efficiency via LLM agents, a revamped development experience that unifies tooling and reduces context switching, and better utilization of digital assets through retrieval‑augmented generation (RAG) and supervised fine‑tuning (SFT).

New Human‑Machine Collaboration Modes

AIGC introduces three roles: the Copilot (assistant for small tasks), the Agent (autonomous multi‑domain expert), and the Facilitator (LLM that integrates information and makes decisions while humans focus on creativity and correction).

Evolution of Knowledge Transmission

Traditional knowledge transfer (talk‑the‑talk, training, mentorship) is being replaced by structured knowledge bases and real‑time model enhancement, enabling rapid assistance for developers through intelligent tools.

Building the Optimal Copilot Collaboration Model

The Tongyi Lingma team focuses on leveraging large models to create the best Copilot experience, emphasizing short‑task handling, human verification to mitigate hallucinations, high‑frequency usage, and short output for efficiency.

Four key developer needs for an ideal Copilot are identified: high‑frequency demand, easy accessibility, accurate intent understanding, and customization for enterprise‑specific contexts.

Model strategies include CodeQwen2 for code completion, Qwen‑plus for specialized coding skills, and Qwen‑max for general R&D Q&A, all integrated via a three‑layer architecture (presentation, core, service) that supports multiple IDEs and ensures data privacy.

Precision Enhancements

Techniques such as trigger‑timing optimization, user‑side self‑learning, adaptive generation granularity, and library‑aware code generation improve accuracy and reduce hallucinations.

Retrieval‑Augmented Generation and Enterprise Personalization

RAG and SFT are applied to enhance code completion and knowledge retrieval, with special attention to local vector storage for security‑sensitive industries.

Future Agent‑Based Software Development

The roadmap envisions a progression from single‑library Q&A systems to coding agents, testing agents, and finally multi‑agent collaboration, aiming to achieve 30‑50% automation on benchmarks like swe‑bench within a year.

Multi‑Agent Conceptual Architecture

A diagram illustrates the planned multi‑agent system, encompassing planning, context handling, tool usage, and environment interaction.

Future Intelligent Development Toolchain

Future tools will blend Copilot and Agent modes, offering AI bots that can operate within IDEs, DevOps portals, or instant‑messaging platforms, providing continuous, context‑aware assistance.

The session concludes with a Q&A where the speaker emphasizes the primacy of technology in product success, the broader potential of multi‑agent systems over Copilot, and current limitations in code optimization.