Iterate on tasks with Copilot CLI autopilot and fleet
Overview. Task: Overview The Copilot CLI has a lot of features! Its documentation is very helpful, but documentation in general doesn’t tell you the patterns of how a tool is actually used that you gain from experience using it. So, rather than me simply reproduce documentation for you (and surely do a questionable job at it), I thought I’d share from my own experience some of the Copilot CLI features I think are actually useful and/or noteworthy to look into, along with a bite-size rundown of how to use each one. ❗Important: turn on /experimental! At the time of this writing, many of these features are experimental. In order you to access them, you must first run /experimental to turn on experimental features. Experimental features are ones that are still in development, but in my experience, they are as stable as any other. 1. Manage the same Copilot session from github.com or your phone with /remote 📜 Docs link: https://docs.github.com/en/copilot/how-tos/copilot-cli/use-copilot-cli/steer-remotely Ever wanted to keep prompting Copilot when you should be attending to real, adult responsibilities instead? Now you can! /remote allows you to interact with the current session either on github.com or on your phone through the GitHub mobile app. Run /remote on to get a github.com link to the session, and then press ctrl/cmd+e to also get a nice little QR code right in your terminal. When scanned with your phone, the QR code will take you straight to the session on the GitHub mobile app. (And no, you can’t access my session by scanning that code. You only have access to your own sessions, so stop that.) Something important to keep in mind is that the session is still running locally from the CLI on your machine. It merely allows you to access that locally-running session from github.com or the mobile app. Practically, this means that your computer needs to stay on. You can use the /keep-alive command to help with this. Also of note is that you’ll quickly notice that slash commands aren’t available on github.com or the mobile app, although skill commands still seem manually invoke correctly. To me, this makes this feature’s use case something along the lines of, “I’m stepping away, but still want to see where things are with Copilot, approve a tool/answer a question it’s asking me, or send a quick correction.” 2. Run deep research with /research 📜 Docs link: https://docs.github.com/en/copilot/concepts/agents/copilot-cli/research Run /research <your topic> to have Copilot research that topic. This can be something about your codebase or something completely unrelated. When it’s finished, it produces a (rather long) markdown document with both its findings and all its sources listed. You can view the research after it’s done with ctrl/cmd+y. In my experience, this usually takes around 10-30 minutes to run, and the research document it produces is around 300-500 lines of content and with 30-50 sources referenced, which makes for quite a thorough document. 3. Share session history and research results as markdown, HTML, or to a gist with /share 📜 Docs link: None This one pairs naturally with /research, since you can share the research results. In addition to sharing research results, you can also share any session’s history as a markdown or HTML file. The different options for this command can be a little confusing at first glance, so here’s the summary: You can export either the current session history or a research document from the current session You can export these as either a markdown file or an HTML file You can export as either a local file or a gist For some example output, here’s what /share html session produced for one of my sessions. It’s quite detailed! 4. Complete a task with multiple subagents in parallel with /fleet 📜 Docs link: https://docs.github.com/en/copilot/concepts/agents/copilot-cli/fleet Running /fleet <prompt> will complete the requested task by spinning up multiple subagents to do as much work in parallel as possible. The subagents used in fleet mode can also be custom agents you’ve created. If you want to be sure your custom agents are used, you can explicitly tell Copilot to use them (and how to use them) in your prompt. According to Evan Boyle of the Copilot CLI team, /fleet works by using a SQLite database for the session that helps inform dependencies between tasks to make the parallelization possible. Pretty cool!; tools: GitHub Copilot. Human role: The original post does not say how a person stayed involved.; limitations: Many features require enabling /experimental mode. /remote sessions depend on the local machine staying on, though /keep-alive helps. /fleet may introduce orchestration overhead, making it slower for small tasks. Autopilot mode risks uncontrolled loops; use --max-autopilot-continues to limit iterations. Source reliability is secondary, based on personal experience rather than official benchmarks..
Tools
Published on Agent Work Atlas:
Overview
Overview
Workflow
Overview The Copilot CLI has a lot of features! Its documentation is very helpful, but documentation in general doesn’t tell you the patterns of how a tool is actually used that you gain from experience using it. So, rather than me simply reproduce documentation for you (and surely do a questionable job at it), I thought I’d share from my own experience some of the Copilot CLI features I think are actually useful and/or noteworthy to look into, along with a bite-size rundown of how to use each one. ❗Important: turn on /experimental! At the time of this writing, many of these features are experimental. In order you to access them, you must first run /experimental to turn on experimental features. Experimental features are ones that are still in development, but in my experience, they are as stable as any other. 1. Manage the same Copilot session from github.com or your phone with /remote 📜 Docs link: https://docs.github.com/en/copilot/how-tos/copilot-cli/use-copilot-cli/steer-remotely Ever wanted to keep prompting Copilot when you should be attending to real, adult responsibilities instead? Now you can! /remote allows you to interact with the current session either on github.com or on your phone through the GitHub mobile app. Run /remote on to get a github.com link to the session, and then press ctrl/cmd+e to also get a nice little QR code right in your terminal. When scanned with your phone, the QR code will take you straight to the session on the GitHub mobile app. (And no, you can’t access my session by scanning that code. You only have access to your own sessions, so stop that.) Something important to keep in mind is that the session is still running locally from the CLI on your machine. It merely allows you to access that locally-running session from github.com or the mobile app. Practically, this means that your computer needs to stay on. You can use the /keep-alive command to help with this. Also of note is that you’ll quickly notice that slash commands aren’t available on github.com or the mobile app, although skill commands still seem manually invoke correctly. To me, this makes this feature’s use case something along the lines of, “I’m stepping away, but still want to see where things are with Copilot, approve a tool/answer a question it’s asking me, or send a quick correction.” 2. Run deep research with /research 📜 Docs link: https://docs.github.com/en/copilot/concepts/agents/copilot-cli/research Run /research <your topic> to have Copilot research that topic. This can be something about your codebase or something completely unrelated. When it’s finished, it produces a (rather long) markdown document with both its findings and all its sources listed. You can view the research after it’s done with ctrl/cmd+y. In my experience, this usually takes around 10-30 minutes to run, and the research document it produces is around 300-500 lines of content and with 30-50 sources referenced, which makes for quite a thorough document. 3. Share session history and research results as markdown, HTML, or to a gist with /share 📜 Docs link: None This one pairs naturally with /research, since you can share the research results. In addition to sharing research results, you can also share any session’s history as a markdown or HTML file. The different options for this command can be a little confusing at first glance, so here’s the summary: You can export either the current session history or a research document from the current session You can export these as either a markdown file or an HTML file You can export as either a local file or a gist For some example output, here’s what /share html session produced for one of my sessions. It’s quite detailed! 4. Complete a task with multiple subagents in parallel with /fleet 📜 Docs link: https://docs.github.com/en/copilot/concepts/agents/copilot-cli/fleet Running /fleet <prompt> will complete the requested task by spinning up multiple subagents to do as much work in parallel as possible. The subagents used in fleet mode can also be custom agents you’ve created. If you want to be sure your custom agents are used, you can explicitly tell Copilot to use them (and how to use them) in your prompt. According to Evan Boyle of the Copilot CLI team, /fleet works by using a SQLite database for the session that helps inform dependencies between tasks to make the parallelization possible. Pretty cool!
Reported result
The source author reports that /research typically takes 10-30 minutes and produces 300-500 line documents with 30-50 sources. No measured business outcomes or success rates are provided. A proposed acceptance check is to verify that a /research output contains at least 30 cited sources and that /fleet completes a parallelizable task without manual intervention errors.
Human role
The original post does not say how a person stayed involved.
Limitations
Many features require enabling /experimental mode. /remote sessions depend on the local machine staying on, though /keep-alive helps. /fleet may introduce orchestration overhead, making it slower for small tasks. Autopilot mode risks uncontrolled loops; use --max-autopilot-continues to limit iterations. Source reliability is secondary, based on personal experience rather than official benchmarks.
Try it
I want to test the GitHub Copilot CLI research and fleet workflow for a technical documentation task. First, ensure I have access to the Copilot CLI and that experimental features are enabled via /experimental. I will provide a topic: [replaceable topic, e.g., 'Best practices for securing Kubernetes clusters']. Step 1: Run /research [topic] and wait for the markdown output. Step 2: Review the generated document for source quality and completeness. Step 3: If the research is valid, use /fleet to generate a summary and a checklist based on the findings. Step 4: Use /autopilot with --max-autopilot-continues 5 to refine the checklist. Please perform a dry run on a small subset first and show me the output before proceeding to the full task. Do not publish or save any files until I confirm the results are accurate. If inputs are missing, ask for clarification on the specific topic or constraints.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.