Microsoft Build 2025 Conference-Open Source GitHub Copilot Chat, launch of the new Copilot Code Agency, full support for MCP...

Frontier Trends
microsoft build conference open source

Microsoft presented a comprehensive strategic blueprint for its move towards the “AI proxy era” and the “Open Agency Web” at its conference in Build in 2025. Microsoft claims that AI’s agent, thanks to the breakthroughs in reasoning and memory, is becoming a central force for development, collaboration and scientific exploration. We envisage a world where intelligence can operate across individuals, organizations, teams, and end-to-end business environments. This emerging Internet vision** is an open network of smart people **, in which artificial intelligence is capable of making decisions and carrying out tasks on behalf of users or organizations.” ** Main highlights: **

Microsoft Strategic Vision:

#1 Reshaping software development process: AI is no longer just “accompanied” but “participant”

##2 Build a powerful, controlled AI proxy ecology

  • Azure AI Foundry Agency online Developer-oriented AI proxy platform to support collaboration between multiple agents (Multi-Agent), context sharing (Model Context Protocol) and semantic Kernel.

  • Observability tools: performance, cost and quality are clearAzure provides a set of proxy monitoring tools to view agent performance, including response speed, reasoning quality, resource consumption, safety issues, etc., and to enhance enterprise control and compliance.

  • Entra Agency ID: Allocating a single identity to each AI agent to avoid a proliferation of “agents” in enterprises and to ensure that the identity, rights and access to data of each AI system are strictly regulated.

#3 # let the company have its own “internal AI employee”

  • Copilot Tuning: Enterprise-defined AI proxy enterprises can quickly train AI agents with internal data (e.g. documents, processes, regulations) without writing codes. For example, firms can build exclusive document-generation agents, styles and expressions can match organizational habits.

  • ** Multi-agent organization: division of tasks, collaborative processing** Multiple agents can be organized into “tasks flow lines” in Copilot Studio, e.g., an analytical document, a generation aggregation, a sending mail to perform complex business processes together.

##4 # Push for Open Agency Web

#5 AI + Scientific research: Microsoft Discory platform

  • ** AI platform for scientific discovery** Microsoft Discory aims to embed “AI proxy” into scientific processes: from data exploration, hypothesis validation, simulation modelling, experimental design, to essay writing, outcome analysis, and process acceleration. It applies to a variety of fields, such as pharmaceuticals, biotechnology, energy, materials, and promotes efficiency gains in basic research and industrial transformation.

Open source GitHub Copilot Chat

Microsoft announced the opening of GitHub Copilot Chat* plugin** and the gradual integration of AI capabilities into the core of the VS Code editor. This marked the formal move of VS Code towards a new phase of “Open Source AI programming environment”.

The background and motivation

The Microsoft team summarized several key current changes in AI development and community feedback as a basis for decision-making to advance open sources: ** Large model capabilities evolved to reduce reliance on the “problem “ ** In the past, in order to achieve better production effects, AI tools often need to use private hints to optimize strategies; but the larger model itself is now more generalized, making this “privatization fine-tuning” strategy no longer a core barrier.

  1. ** User experience (UX) mode matured** AI interactions in different editors (such as chat windows, automatic code completion styles, etc.) are increasingly standardized. Microsoft wants to promote open sharing of these functions through open-source approaches, encouraging communities to continue grinding.
  2. AI plugin ecological expansion The developers want to build their own AI plugins, but because the Copilot Chat plugins are closed-sourced, debugging and acceptance limits are higher. When open-sources, developers can learn more easily to achieve logic and achieve depth integration.
  3. Increased transparency in the use of data Community users are concerned about what VS Code AI plugins collect, and with open source plugins, all data-processing logic will become clear and help build confidence.
  4. Improving safety and using communities to detect loopholes History has shown that open sources help to detect and repair security gaps more quickly. This open mechanism is particularly important for AI editors.

# Technology programmes and follow-up planning

Copilot Chat plugin open source**

  • The plugin code will use MIT License authorization;

  • Core functions will be “dismantled and integrated” in the VS Code core module over time after open source;

  • All contributors can participate in the optimization and re-engineering of the plugin.

** The Hint test frame will also be open**

  • In order to address the “non-determinate question” of the results of the large model reasoning, Microsoft will open up the internal use of the prompt test framework;

  • This will help the contributor verify whether the function is stable and the output is manageable when submitting Pull Request.

** Developer contribution path will be simplified**

  • AI related functions will be as easy to contribute as other VS Code core functions;

  • Future contribution prompt, UI logic, model access logic will all have uniform criteria.

Version management and progress tracking

  • All developments will be updated on a monthly basis in the Development Plan at VS Code;

  • The official FAQ page also provides simultaneous answers to frequently asked questions from the community.

# The real value after the open source

Meaning to the developer The plugin development can learn the plugin source, reuse the generic components, better debugger the access custom model no longer relies on the official Copilot, can be replaced by other LLM security and data, identify which data will be collected and used, increase trust, contribute the threshold, lower learning curves, more complete testing infrastructure support The meaning of the whole AI programming ecology.

  • To provide a stable open-source infrastructure for standardized development;

  • Encourage global developers to collaborate in the formation of diversified plugins and model-adaptation layers;

  • Promote the new era of AI programming from “product trials” to “Developers’ custom-based tool chains”.

New programming agent: Copilot Coding Agency

同时微软发布了新的Copilot Coding Agent，一个可被分配任务、自动提交代码、并集成 CI/CD 流程的 AI 代理，具备自主编码、分析和迭代的能力。 It is no longer just an auxiliary writing code, but rather an “AI Agent” that can take the initiative** to perform development tasks**.

Core competencies and workflows

  1. ** Job-driven automatic development**

-Agent will automatically: Activate Secure Virtual Environments (based on GitHub Actions)

  • Cloning the code library, parsing the context (using enhanced retrieval + RAG technology)

  • Write, push code to Drawr PR

  • Update PR Description and Record Behavior Log

  1. ** Capability to understand context**

  • Analysis of mission intent using GitHub code search, file structure and historical discussion

  • Support the extraction of pictures from PR/Issue for visual analysis (e. g. UI Mocup)

  • Respond to the Human Developer’s comments on Draft PR and continue over time

  1. ** Multi-model call capability (MCP protocol support)**

  • Through Model Context Protocol (MCP), Agent has access to non-GitHub data sources (e. g. intra-enterprise API or external document database)

  • User can configure custom MPCP Server in project settings

Security and access control mechanisms

In order to ensure the security of the team code library, Copilot Agent designed a set of strict protection strategies:

Use the scene and the applicable boundary

• Applicable task type:

  • Low-moderate complexity, clear needs, clear borders Bug repair, unit test completion, code re-engineering, document update, etc.

  • Initial realization needs template, migration code structure, split logic module, etc.

Not applicable:

  • Highly innovative, lacking clear context, and highly dependent on cross-domain tasks

  • Needs requiring a great deal of subjective judgement and in-depth product understanding

Integrating approach to environmental support

  • Default operating environment: GitHub Actions

  • Local Agent model support: VS Code, Xcode, Eclipse, Jet Brains, Visual Studio, etc.

  • Enterprise users can set the activation policy uniformly and control which items will enable the Agent mode

Use policy and billing instructions

The difference between the traditional Copilot 更多Microsoft Build 2025 大会内容：https://blogs.windows.com/windowsdeveloper/2025/05/19/advancing-windows-for-ai-development-new-platform-capabilities-and-tools-introduced-at-build-2025/

← Previous Google releases Gemini Diffusion's diffusion-based model at a speed of 2000 token/seconds comparable to Gemini 2.0 Flash-Lite
Next → Gemini 2.5 Pro details recent developments in video understanding that can handle about 6 hours of video with a maximum support of 2 million token context

Related Posts

Microsoft releases Bing Video Creator driven by Sora to generate video free of charge
microsoft
Microsoft developed an AI doctor for MAI-DxO who can ask questions, open tests, analyze the results, and eventually determine the cause of the disease, which is 85% accurate.
microsoft
Kyutai team open-source real-time voice model Kyutai STT listens to text and you're smart enough to judge if you're done talking.
open source
MiniMax releases the MID open source model MiniMax-M1 training costs only $5.35 million.
open source