AI at Work
Assigning Work to GitHub Copilot's Cloud Agent: Define the PR and Check Workflow Approvals
A complete fictional dependency issue and mixed revision-specific PR review record distinguish agent work, privileged CI execution, independent required review and merge/release decisions; no account test.
Before assigning work to GitHub Copilot's cloud agent, define the change you want reviewed, the files and permissions it must leave alone, and the evidence reviewers need before accepting the PR. Then check the repository's actual workflow-approval policy. Permission to prepare a branch is different from permission to execute privileged CI, approve a contribution, merge it, or deploy it.
This guide checks GitHub's public documentation as of October 25, 2026. It belongs to LandOffer's workplace content, and uses a fictional dependency-maintenance task to demonstrate the handoff. No repository was assigned, agent started, workflow approved, PR created, or package installed for this article. All case records below are authored examples, not a Copilot reliability test.

Confirm the execution surface and the requested deliverable
GitHub's current Copilot on GitHub.com documentation describes cloud-agent work in an ephemeral cloud development environment. It can research, plan, edit files, and run tests. A current conversation can progress from planning to implementation; a PR can be requested at the start or created after reviewing changes. Do not assume every research question immediately opens one.
That is a different execution location from an agent editing your local IDE workspace. GitHub also documents delegating from an IDE to the cloud agent. In the VS Code route, including local changes can push them so the cloud session starts from that state. An IDE entry point therefore does not mean the task stays local. A remotely controlled CLI session is another surface; its machine and permission context must be checked separately.
Before proceeding, identify the repository, base revision, account access, current organization policy, execution environment, and desired result: research, a plan, a branch, or a draft PR. The fictional task here explicitly requests a draft PR in one repository. It does not authorize publishing a package or changing another repository that consumes it.
Treat the assignment description as a bounded work agreement. Include the behavior to preserve, not just a version bump. A small maintenance request can become a large unreviewable change if the agent interprets a failed install as permission to replace the build system or weaken its checks.
Write a complete issue with limits and acceptance criteria
The following issue is wholly fictional. note-format-core, its versions, filenames, and test outcomes are invented for this example; they are not a recommendation about a real dependency or a reported vulnerability. The imaginary repository builds internal engineering notes. Maintainer Hana wants a narrow patch upgrade, not a formatting redesign.
Issue M-42: Prepare a draft PR for the note formatter patch
| Field | Agreed issue content |
|---|---|
| Change | Update note-format-core from 2.6.3 to 2.6.4 in tools/notes/package.json and its existing lockfile |
| Reason | Evaluate the patch against our current formatting contract; make no security-fix claim without an advisory |
| Starting point | Repository engineering-notes; owner-selected base revision B0 recorded in the issue |
| Allowed implementation | Minimal dependency and lockfile update; focused compatibility test in tools/notes/tests/format.test.ts if needed |
| Preserved behavior | Heading order, blank lines, literal code-block content, and the final newline must match accepted fixtures |
| Forbidden changes | No .github/workflows/, agent setup, runner labels, permissions, secrets, registry configuration, production credentials, release scripts, or unrelated dependency upgrades |
| Validation requested | Existing ten formatting fixtures, package type check, and the current documentation build; report exact commands, revision, results, and any omitted check |
| Stop condition | Stop and explain if access is missing, lockfile churn crosses unrelated packages, tests require a new permission, or behavior changes beyond the patch |
| Deliverable | One draft PR with changed-file rationale, dependency/lockfile explanation, results, failures, and unanswered questions |
| Human owners | Hana owns task scope; platform owner Malik reviews execution permissions; independent maintainer Orla owns required code review |
| End boundary | Do not merge, publish, deploy, disable a check, or expand access to finish the task |
These limits give the agent a concrete reason to stop and escalate. If the dependency requires a private registry credential, “complete the update” does not resolve who may grant that access. Hana must route the blocker to the authorized owner rather than paste a token into the issue.
The fixture requirement also makes a later result assessable. “Tests pass” is too broad if the package suite contains ten cases but the agent ran one selected test. A useful handoff preserves failures and omissions so reviewers can decide the next step. It becomes misleading when a missing check disappears behind a confident PR summary.
For an actual task, attach only approved context. A minimal reproduction or sanitized expected-output file is often enough. Do not include a customer document, access token, or production dump merely because the agent needs an example of formatting behavior.
Inspect access, runners, and workflow approval separately
GitHub's guardrails guidance recommends planning agent policies, protecting sensitive configuration, and reviewing minimum workflow permissions. It recommends fresh GitHub-hosted runners, or ephemeral runners when self-hosting. These controls belong with repository and organization administrators; a prose instruction does not enforce them.
The runner configuration documentation describes an organization default and possible repository overrides. A runner trusted to access internal resources deserves a separate decision about what an agent can reach. Check the effective runner group and setup configuration, not just the label “cloud.” A fresh environment also does not prove that every tool or dependency inside it is safe.
Keep secret types distinct. The current Agents secrets documentation describes dedicated Agents secrets and variables exposed to the agent environment. It says ordinary Actions, Codespaces, and Dependabot secrets are not passed to cloud-agent sessions. Separately inspect the sensitive access available to a later PR workflow. Inventory both execution paths with the authorized owner.
GitHub's agent settings guide says PR workflows do not run automatically by default when Copilot pushes changes. It also allows repository administrators to disable the approval requirement. Therefore, check the actual setting before assigning the task; do not present default blocking as an unconditional guarantee.
The guide warns that workflows can hold sensitive privileges and recommends inspecting proposed changes, especially workflow files, before using Approve and run workflows. That button authorizes execution. It does not certify the diff or approve a merge. In M-42, Malik must inspect scripts invoked by CI as well as workflow YAML: a changed package script can alter execution even when the workflow file is untouched.
Review a proposed PR using evidence tied to its revision
Here is a completed fictional review record with a deliberately mixed outcome. Revisions A1 and A2 are narrative labels, not real commit hashes. In an actual PR, replace them with verifiable revision identifiers and link the relevant diff and logs. No commands or tests in this example were executed.
The imagined first PR at A1 updates the target dependency, but also edits .github/workflows/notes.yml to omit a failing build. Hana rejects that scope expansion. Her proposed follow-up is specific: remove the workflow edit, keep the existing check, and explain the build failure. This is an example of a review message, not a comment actually sent to Copilot.
At A2, the imagined agent removes the workflow edit and reports that nine of ten formatter cases passed in its session; the Unicode heading fixture failed. It reports a successful type check and a documentation build blocked by unavailable registry access. The review record preserves those statements as unverified agent claims until their underlying evidence is inspected.
| Review item at A2 | Supplied example evidence | Human assessment and next owner |
|---|---|---|
| Changed files | Manifest, lockfile, focused test; workflow edit absent from A2 diff | Hana accepts the visible scope correction only; compare the full diff with B0 |
| Dependency resolution | PR summary says only the target package changed | Orla needs the lockfile comparison; summary alone is insufficient |
| Formatting contract | Claimed nine passes and one Unicode failure | Compatibility criterion unmet; Hana requests the exact input, expected output, and failure log |
| Type check | Agent reports success | Require command, exit status, and log tied to A2; not equivalent to complete validation |
| Documentation build | Reported blocked by registry access | Malik decides whether an approved validation route exists; no new credential granted |
| Workflow execution | Approval still pending in this fictional record | Malik has not authorized PR CI; no PR workflow result claimed |
| Required review and merge | Orla has not approved; task requester Hana is involved | Required review unresolved; keep PR unmerged |
The resulting decision is request changes; do not approve workflow execution or merge yet. A2 fixes one scope violation but fails a behavior condition and lacks a build result. A patch version number does not resolve the Unicode failure or the missing documentation build.

Route each next action to the person authorized to decide
The case can move forward without relaxing its boundaries. Hana asks for the Unicode failure evidence and decides whether the proposed patch is still the right task. Orla checks the lockfile and compatibility change. Malik checks the approved runner, invoked scripts, and registry-access options. Each owner records a decision or a concrete blocker; “waiting for approval” should identify which approval and why.
If Malik later authorizes PR workflows after inspecting a new revision, record that revision and inspect the resulting run. A queued run is not a passed check. A completed run needs its conclusion, required jobs, and skipped or cancelled steps reviewed. Fresh commits require a fresh look at what will execute and which evidence still describes the current diff.
GitHub's output review guidance says the person who requested a Copilot PR cannot supply an approval that counts toward required PR approvals; another reviewer is needed. In this fictional team, Orla fills that independent role. Malik's authority to allow CI does not substitute for her code review.
The risk and mitigation documentation says the cloud agent cannot approve or merge a PR and describes branch restrictions and human review controls. Maintainers still judge correctness, sensitive access, and readiness within those built-in protections. A verified agent-authored commit is provenance, not a passing acceptance test.
M-42 stays open at the end of this record. There is no invented successful rerun, approval, or deployment. If an authorized person later merges an accepted PR, release ownership and any deployment gate still need their own explicit decision under the team's process.
Finish with a reviewable handoff rather than a broad success claim
A useful final handoff identifies the base and current revision, explains every changed file, links actual check evidence, and lists failures and blocked work. It also states who must decide next. For M-42, that means a corrected scope, an unresolved Unicode failure, a missing build, pending workflow authorization, and independent review still outstanding.
Before assigning your own task, fill those fields and confirm the effective repository policy. Before accepting its result, compare the current diff, required runs, and unresolved failures with the issue's conditions. Keep task assignment, workflow execution, required PR review, merge, and release decisions separate. LandOffer's workplace content provides related collaboration guidance; the evidence for accepting this code change must come from your repository and authorized reviewers.
Sources and Further Reading
- GitHub Docs: Copilot on GitHub.com.
- GitHub Docs: Use cloud agent in your IDE.
- GitHub Docs: Build cloud-agent guardrails.
- GitHub Docs: Configure agent runners.
- GitHub Docs: Agents secrets and variables.
- GitHub Docs: Configure agent settings and workflow approval.
- GitHub Docs: Review Copilot output.
- GitHub Docs: Risks and mitigations.


