What Is GitHub Copilot Cloud Agent?¶

For / Key Points

For: Developers and technical leads who want to delegate a GitHub issue and receive a reviewable change

Key Points:

Copilot coding agent became Copilot cloud agent in April 2026, with a workflow that no longer requires an immediate pull request
IDE Agent Mode works synchronously in your local environment; cloud agent runs asynchronously in a GitHub Actions environment
Design tasks around one repository, one branch, one pull request, and a hard 59-minute session limit

Update the product name first¶

On April 1, 2026, GitHub reframed Copilot coding agent as Copilot cloud agent. The older workflow began with a pull request. The current agent can research a repository, write an implementation plan, change a branch, show a diff, iterate, and create a pull request only when the work is ready.¹

Older configuration and metrics may still use coding agent. This guide uses the current cloud-agent name.

Cloud agent versus Agent Mode¶

Both can inspect code, run commands, and edit multiple files. The practical difference is the execution environment and whether a person must remain in the loop while work runs.²

Dimension	Copilot cloud agent	IDE Agent Mode
Execution	Ephemeral GitHub Actions environment	Local IDE and workstation
Interaction	Asynchronous, in the background	Synchronous conversation
Output	Branch, diff, and optional pull request	Local working tree
Best fit	Bounded issues, routine maintenance, test additions	Exploratory debugging, design discussion, local dependencies
Human role	Review the result later	Guide the work as it happens

Choose based on supervision, not a vague performance ranking. If acceptance criteria fit in an issue, cloud agent can work well. If the task requires repeated product or design judgment, use Agent Mode with a person present.

The current workflow¶

flowchart LR
    A[Issue or prompt] --> B[Research and plan]
    B --> C[Change an isolated branch]
    C --> D[Test and self-correct]
    D --> E[Review the diff]
    E -->|More instructions| C
    E -->|Ready| F[Create a pull request]
    F --> G[Human review]

You can ask for a pull request immediately, or keep iterating on the branch before opening one. Asking for a plan first is useful when an issue is large enough to contain several possible implementations.

Tasks that fit¶

Small bugs with reproduction steps and expected behavior
Incremental features that follow an existing pattern
Missing tests for a clearly bounded module
Documentation, typing, and lint fixes
Dependency or technical-debt work with mechanical acceptance checks
Simple merge-conflict resolution inside one repository

Split or keep human-led any task that needs multiple repositories, exceeds the 59-minute execution window, touches production data or authorization boundaries, depends on local-only infrastructure, or has subjective acceptance criteria.

The important boundary is not lines changed. It is whether completion can be verified.

Put five things in the task¶

## Goal
Fix the duplicate error notification after a failed login.

## Scope
- Change only `src/auth/`.
- Do not change the API response shape.

## Reproduction
1. Submit an invalid password.
2. Two notifications appear.

## Acceptance criteria
- Exactly one notification appears.
- Existing authentication tests pass.
- Add one regression test.

## Checks
`pnpm lint && pnpm test auth`

Goal, scope, reproduction, acceptance criteria, and commands give the agent a plan and a definition of done.

Prepare the environment before delegating¶

Cloud agent runs in an ephemeral GitHub Actions environment. Define missing dependencies and setup in .github/workflows/copilot-setup-steps.yml.³

Keep setup limited to public build dependencies. Never write long-lived credentials into repository files. Pass required secrets through GitHub's controls with the narrowest practical permissions. Setup time also consumes the 59-minute session, so cache design and a fast validation path matter.

Current limitations¶

GitHub documents these material boundaries:²

A task changes only the selected repository
It works on one branch and can open exactly one pull request
A session has a hard 59-minute maximum
The repository must be hosted on GitHub
Some rulesets or branch-protection rules are incompatible unless configured for Copilot
Copilot cloud agent does not honor Copilot content exclusions

The last limitation is a security boundary. Do not treat content exclusion as isolation from cloud agent. Decide separately which repositories and data the agent may access.

Plans, Actions minutes, and AI credits¶

Cloud agent is available with paid Copilot plans, subject to organizational policy.² Its sessions consume GitHub Actions minutes and AI credits. AI-credit usage depends on the selected model and tokens processed.

Do not default every task to the most expensive model. Use faster models for routine transformations and higher-reasoning models where failed attempts would cost more than the model choice.

Review the pull request as a human-owned change¶

Passing tests do not make a change automatically acceptable. Check:

Whether every acceptance criterion is satisfied
Whether the diff stayed inside the requested scope
Whether dependencies, permissions, or outbound connections changed
Whether tests cover failure behavior, not only the happy path
Whether rollback is practical
Whether CI reproduces the checks the agent reported

Copilot code review can provide a first pass, but two agents from one vendor can share blind spots. Humans still own security boundaries, business correctness, and product judgment.

Start with a narrow rollout¶

Select one repository and three documentation or test tasks
Record acceptance rate, return reasons, Actions minutes, and AI credits
Improve AGENTS.md and environment setup
Expand to bounded bug fixes
Keep separate approvals for high-risk areas

Measure adopted changes and review time, not generated code volume.

Summary¶

Copilot cloud agent is not simply Agent Mode hosted elsewhere. It is an asynchronous GitHub worker that can research, plan, change, and iterate before a pull request exists.

Remember three boundaries:

One task, one repository, one branch, one pull request
A maximum of 59 minutes
Content exclusion is not an isolation control for cloud agent

Tasks shaped to those boundaries can gain asynchronous throughput without sacrificing reviewability.