Custom Copilot Agents: GitHub Seeks Feedback on Partner Tools

GitHub has moved its custom agents for Copilot from conference announcement to real product. It launched a collection of partner-built agents and opened a public channel for developers to say what works and what doesn't. The launch is ambitious. The feedback loop is just getting started, and how much friction users hit in practice remains an open question.

What GitHub shipped

On December 1, 2025, GitHub posted a "feedback wanted" announcement in its community forum's Copilot News and Announcements category. The post described custom agents as first unveiled at GitHub Universe and now expanding with partner-built options 1. Two days later, a GitHub blog post titled "Your stack, your rules" formally introduced the agents. It pitched them for debugging, securing, and automating engineering workflows across the terminal, the editor, and github.com 2.

The partner categories are broad. GitHub listed observability, infrastructure as code, security, databases, incident response, experimentation, and automation 1. The company frames these agents as tools for running "real engineering workflows" rather than simply generating code 1.

How custom agents work

The design is deliberately lightweight. GitHub describes custom agents as domain experts defined in Markdown that extend the Copilot coding agent. They need no infrastructure of their own 1. They are meant to run anywhere Copilot does [1]:

• the Copilot CLI in a terminal
• Copilot Chat in VS Code
• the Copilot panel's Agents view on github.com

Example tasks include scanning dependencies and optimizing other parts of the development pipeline 1. The forum post also separates partner-built agents from community-built ones. That suggests GitHub wants an ecosystem broader than a curated vendor list 1.

Where this fits in Copilot's trajectory

The launch is the latest step in a year-long push to turn Copilot from an autocomplete tool into an agent platform. GitHub's product feed shows the progression [2]:

• April 2025: Agent Mode with Model Context Protocol (MCP) support rolled out to all VS Code users. The same announcement covered a Copilot Pro+ tier with premium requests, general availability of models from Anthropic, Google, and OpenAI, next edit suggestions, and a Copilot code review agent.
• August 2025: An Agents panel let developers hand off coding tasks from anywhere on GitHub. Copilot works in the background, opens a pull request, and tags the user for review.
• September 2025: A new embedding model was billed as making code search in VS Code faster, lighter on memory, and more accurate.

Custom agents build on that foundation. If the coding agent is the engine, Markdown-defined agents are a way to give it specialized instructions for a particular tool or domain. Partners supply that domain knowledge, and GitHub supplies the runtime.

Reading the feedback request

Framing the community post as a feedback solicitation is telling. GitHub is asking developers to shape a product that is already live in several surfaces. At the time captured, the thread had drawn only three comments and two replies 1. That is too few to support broad claims about developer sentiment in either direction.

Some questions follow naturally from the design, though they are analysis rather than reported complaints. Markdown-defined agents are easy to write. They may be harder to reason about once several are active and a developer has to predict which agent handles which part of a workflow. And because the agents extend the Copilot coding agent, which does its work in the background and returns pull requests 2, they likely inherit whatever constraints apply to that agent's execution environment. That could include limits on how they interact with existing build and CI pipelines. Whether those concerns turn out to be real is exactly what a public feedback thread should reveal.

Why it matters

The strategic bet is that development teams want AI help across their whole toolchain, not just in the code editor. GitHub has recruited partners in observability, security, and infrastructure 1. It is positioning Copilot as a coordination layer for operational work that has traditionally lived in separate dashboards and consoles.

The zero-infrastructure, Markdown-first approach lowers the barrier for both vendors and individual developers to publish agents 1. That could speed up the ecosystem. It also shifts work onto users to judge quality, scope, and how an agent behaves inside their own pipelines.

GitHub is setting the direction, and its partners are providing the breadth. Whether custom agents become routine or remain a novelty will depend on how quickly GitHub turns early feedback into clearer workflow guidance and fewer edge-case surprises. On current evidence, the platform is shipped and the verdict is still pending.