This repository contains an example of GitHub Copilot handoffs for agents. The goal of this implementation is to have a simple, clean experiment to understand the inputs and outputs.
This repository is part of a series of repositories exploring GitHub Copilot-specific features. To find other repositories, search for topic:0-awesome-copilot.
Using number-converter-first agent, provide a number: 1234 as input.
This should trigger the whole flow and produce the final decorated output:
| N | R | E | S |
|---|---|---|---|
| 1234 | MCCXXXIV | 1️⃣2️⃣3️⃣4️⃣ | one thousand two hundred thirty four |
This implementation uses a SKILL.md file to keep the critical "Do not stop working on this issue until there are no further subagents to run." instruction.
As of 12/25/2025, this feature requires an experimental flag to be enabled in your GitHub Copilot settings: chat.useAgentSkills. It seems to work with every model except the gpt- models.
Works well with:
- Grok Code Fast 1
- Raptor mini (Preview)
- Claude Haiku 4.5
- Claude Opus 4.5
- Gemini 3 Flash (Preview)
- Gemini 3 Pro (Preview)
Does not work with:
- GPT-5.2
- GPT-5.1-Codex
- GPT-5.1-Codex-Max
This isn't currently documented, so your mileage may vary. You can test it by doing the following:
Select agent-number-converter-first and provide a number as input. If the agent can call the subagents and produce the final decorated output, then it's working. Try it with Grok Code Fast 1.
If you select a non-compatible model, the agent will just return the output of the first subagent and stop there.
Created an issue reporting this observation: #285075
General observations live in a separate repository: MarioRicalde/copilot-observations.