What changes: the task is split in two
Autocomplete finishes the line you are typing. An agent takes on a whole task. In the Copilot cloud agent model, you write the task and the agent works in the background. You review what it produces later.
Writing the task and reviewing the result no longer have to happen at the same time or in the same place. That split is where the flexibility comes from.
What Copilot cloud agent can do
GitHub's documentation lists many ways to start a session. Based on the GitHub docs in the evidence packet is not the right basis here, so these are the documented entry points as GitHub describes them:
- On GitHub: issues, the agents tab, the dashboard, Copilot Chat, new repositories, and failing GitHub Actions runs.
- On mobile: GitHub Mobile.
- In IDEs: Visual Studio Code, JetBrains, Eclipse, and Visual Studio 2026.
- Through tooling and integrations: the REST API, GitHub CLI, the GitHub MCP Server, Jira, Slack, Microsoft Teams, Azure Boards, Linear, and Raycast.
Sessions can also start automatically, on a schedule or when an event happens, such as an issue being opened. Copilot works in the background. Some entry points open a pull request on their own. In other cases you ask Copilot to open one from the session logs.
A plausible workflow is to assign a well-specified bug from your phone or from Teams, then inspect the session and the pull request at a workstation. The documentation does not promise a turnaround time, a correct fix, or passing tests. Treat the agent's output as a proposal.
Who can use it
"Every paid plan" is only part of the picture. GitHub's documentation sets these boundaries:
- The agent is available on all paid Copilot plans.
- It is not available in repositories owned by managed user accounts. GitHub explains that it runs on GitHub-hosted runners, which those personal repositories don't have.
- It can be explicitly disabled for a repository.
- For organization repositories, organization or enterprise administrators control availability. For Business and Enterprise plans, an owner must enable the Copilot cloud policy.
- Copilot responds to pull request comments only from people with write access, and only on open pull requests.
GitHub's troubleshooting guidance also covers common failure points:
- Sessions that stall: a session on GitHub Actions that stays stuck times out after an hour. You can retry by unassigning the issue and reassigning it to Copilot.
- Workflows that don't run: GitHub Actions workflows don't run automatically when Copilot pushes. A person has to click the approve-and-run button in the pull request.
- Chat in IDEs: in VS Code, Visual Studio, and JetBrains, you must mention the @github chat participant to create a pull request from chat.
- CI failures: GitHub says Copilot is more likely to avoid pushing changes that fail CI if you give it clear instructions in a .github/copilot-instructions.md file.
- Network access: a firewall limits Copilot's internet access by default, to reduce the risk of data leaking.
- Screenshots: attached images over 3.00 MiB are removed from the request.
How many developers use agents
The adoption numbers need careful framing. Stack Overflow's pulse survey from late April 2026 had 1,100 respondents. Stack Overflow reported that 59% use agents at work at any frequency, up from 31% in the 2025 Developer Survey. That is a survey of respondents, not a census of all developers.
The same survey shows the use is cautious:
- 60% of respondents block agents from making unapproved system changes.
- 68% prefer single-agent setups over multi-agent ones.
- 63% rarely or never let agents run fully on autopilot.
- GitHub Copilot (65%) and Claude Code (50%) were the single-agent workflows of choice among full-stack developers.
So the flexibility is real, but most people using it keep a human in the loop.
Remote work figures
Nerdbot quoted an "82% have some remote flexibility" statistic and a trend table. I could not verify either against Stack Overflow's 2026 data, so I'm leaving them out. The 2026 Developer Survey work-setting data has 24,602 respondents, and it splits them this way:
| Work setting | Share |
|---|---|
| Remote | 34.8% |
| Hybrid, leaning in-person | 19.9% |
| In-person | 17.6% |
| Hybrid, leaning flexible | 15.5% |
| Very flexible, your choice | 12.2% |
Nothing here shows that agents caused any shift in work arrangements. What it does show is that most respondents have some choice over location. That fits asynchronous delegation, but it isn't proof of it.
Flexibility is not speed
The METR randomized trial is the most-cited study here, and it is often misquoted. It covered 16 experienced open-source developers working on 246 tasks in familiar repositories, mostly with Cursor Pro and Claude 3.5 and 3.7 Sonnet. METR's own update says the early-2025 study found tasks took 19% longer with AI, with a confidence interval of +2% to +39%. Developers had expected a 24% speedup and still believed afterward that they had been about 20% faster.
The finding is old. METR has since said that finding is outdated. It also said it believes developers are likely more sped up in early 2026 than in early 2025. But METR called the data from its follow-up experiment an unreliable signal, because many developers declined to take part if they would have to work without AI. That likely biases the measured speedup downward. Neither number is a reliable forecast for your team.
Two points hold up either way:
- Perception is a poor gauge. The developers' own sense of speed was off by a wide margin in 2025.
- Measure your own team. Compare cycle time, defect rate, and review time before and after adopting agents.
The hidden cost is review
Ten agent pull requests a day means ten reviews. If those reviews pile up, the time the agent saved goes to waiting. This is general engineering reasoning, not a measured finding, but it is the main risk of the queue-it-and-leave workflow.
Practical habits
These are sensible controls to consider, not tested results:
- Keep tasks small. Aim for one bug or one feature slice per task.
- Write specific issues. Name the file or module, the expected behavior, the test that should pass, and any limits, such as not changing the public API.
- Require tests, but don't trust them blindly. Passing tests can still miss the intent of the task.
- Limit permissions. Keep agents on branches, keep them from merging on their own, and keep secrets out of their environments.
- Book review blocks. Fixed daily review time keeps output from sitting for days.
- Keep an instruction file in the repository. For Copilot, GitHub points to .github/copilot-instructions.md for build commands, style rules, and test steps.
Bottom line
Cloud agents make some coding work asynchronous, and Copilot's documented entry points reach into tools people already use, including Teams, mobile, and the IDE. Whether that makes a team faster depends on how well the tasks are written, how fast reviews happen, and what the team actually measures. Plan eligibility and admin policy also decide whether you can use it at all.
References
- How Developers Are Gaining Work Flexibility With Agentic AI - Nerdbot Nerdbot · 2026-11-20T14:39:22+00:00
- Starting GitHub Copilot sessions - GitHub Docs docs.github.com
- Troubleshooting GitHub Copilot cloud agent - GitHub Docs docs.github.com