# AI-Assisted OTA Releases: What Should a Human Approve?

> Review an AI-prepared React Native OTA release using exact artifact identity, compatibility tests, rollout scope, and recovery evidence.

Canonical: https://bundledrop.app/resources/ai-agent-ota-release-approval

Published: 2026-11-14

Updated: 2026-11-14

Category: Guides

Topics: OTA Operations, Rollouts & Recovery

An AI agent can inspect a change, run checks, prepare release notes, and collect the artifacts for a React Native OTA release. The human approving it still needs to know exactly what will reach devices, why it is compatible, and how the team will respond if it fails.

A useful approval request identifies one candidate, links its test evidence, and describes one bounded change to release state. “The agent says it is ready” leaves too much for the reviewer to reconstruct.

This guide focuses on that handoff. The [OTA CI/CD guide](/resources/react-native-ota-ci-cd) covers the pipeline that produces and retains the bundle. Here, the question is what evidence lets a person approve the next release action.

## Decide where preparation stops

Before the agent runs an upload, establish which project and channel it may change and whether that action can expose code to users.

In Bundle Drop, [managed publishing](/docs/managed-publishing) is configured per channel. With it enabled, a compatible bundle needs an active or completed rollout to become eligible. Without it, the newest compatible, non-revoked bundle can become eligible after upload. An upload into a channel called `production` is not automatically a review-only step.

Choose the boundary explicitly: prepare locally, upload to an approved test channel, or upload behind an already verified publishing gate. Confirm the actual channel setting rather than inferring it from the channel name. Changes to that setting need their own review because they change how saved rollout state affects delivery.

Keep the automation's credentials within the permissions needed for the assigned work. If the available credential also permits production changes, a prompt saying “prepare only” does not remove that access. Keep production credentials out of preparation jobs where the workflow allows it, and give the reviewer a separate release step.

## Tie the evidence to one artifact

A source commit explains where the change came from. It does not identify everything that went into the uploaded bundle. Dependency resolution, platform, build inputs, and generated assets also matter.

The approval record should connect:

- the source revision, any uncommitted changes, and the reviewed diff;
- the locked dependencies, SDK version, and CI run or local build record;
- the organization, project, API destination, and channel;
- the platform, native app version, and runtime compatibility value;
- the uploaded bundle identity, version, and full hash;
- the retained bundle, matching source map, and test results.

Use the exact hash in the record. A display label such as “latest Android” can refer to a different candidate by the time someone approves it.

Bundle Drop's [CI/CD result file](/docs/ci-cd#artifacts-and-the-result-file) supplies upload metadata when `--artifact-dir` is used. Request a source map with `--sourcemap`, then verify that the retained files exist and belong to the candidate. The result file is one input to the review; it does not contain the code review or prove that QA exercised the release.

Check the installed CLI's result contract rather than assuming every SDK version returns the same fields. Keep credentials out of the approval record and restrict access to source maps and build logs according to the application's security requirements.

## Show compatibility and the limits of testing

Ask the agent to classify the change before summarizing the tests. A JavaScript fix that calls a newly added native method still depends on a new binary. The [OTA-versus-store decision guide](/resources/when-ota-update-isnt-enough-store-release) owns that classification; the approval record should contain the conclusion and the evidence relevant to this change.

Identify the native release builds used for testing and the platform/runtime lanes they represent. If native fingerprinting is enabled, retain the matching build inputs and any required build receipt. A fingerprint match constrains native compatibility; it does not establish that the application behaves correctly. See the [native fingerprinting documentation](/docs/native-fingerprinting) for its current configuration and limits.

Report checks as passed, failed, or not run. Include failures that were fixed, followed by the rerun on the final candidate. A passing test from an earlier bundle should not silently become evidence for a rebuilt one.

For a login fix, useful evidence might include the original reproduction, a successful login, an expired session, an interrupted network request, and a cold launch after applying the identified OTA bundle. The exact cases depend on the change. A broad test count without those connections gives the reviewer less information.

State the gaps plainly. “Android release build tested; iOS not run” is actionable. “All checks passed” is misleading if it hides an untested platform. Narrow the proposed release to the supported evidence or explain why additional testing is required.

## Make the approval request describe the change

The reviewer should be able to compare the current release state with the proposed state without interpreting an agent's command history.

Include the destination channel, platform/runtime lane, candidate hash, current rollout state, proposed percentage, and audience rules. Identify any promotion needed before the rollout. Bundle Drop [channel promotion](/docs/channels#promoting-bundles-to-another-channel) reuses an existing bundle, but the destination's history and publishing controls still determine eligibility.

An approval request can be short while remaining precise. For example, using illustrative values:

> Approve setting an active rollout at 10% for the already promoted Android bundle v42, full hash recorded in the linked evidence, in the Acme / Shop project's production channel for runtime 4.0.9, with targeting disabled? Managed publishing is enabled and this candidate has no active rollout. The attached release-build tests passed. iOS is outside this request. Maya will review the agreed health signals before any increase.

The actual record needs the full hash and working evidence links, not placeholders. If promotion is still required, include that action and its destination explicitly rather than treating it as an invisible prerequisite.

Record approval against that candidate and those settings. Re-read the channel and rollout state immediately before execution. Another operator may have changed them while the review was open. If the candidate, destination, audience, or proposed action changes materially, return the changed proposal for review. A fresh read reduces stale decisions; it does not provide an atomic lock against concurrent changes.

## Agree on observation and recovery before exposure

The agent should prepare the observation plan alongside the release request. Name the responsible engineer, the metrics, the baseline, the observation window, and the stop conditions. Choose thresholds from the application's normal behavior rather than borrowing a universal crash-rate target.

Keep configured exposure separate from observed adoption. A 10% rollout describes eligibility, not proof that 10% of devices are running the candidate. Devices still need to check, download, and activate under the app's update policy. Identify the running OTA artifact in monitoring events so failures can be attributed to this release; the [source-map guide](/resources/react-native-ota-sentry-source-maps) explains that identity chain.

Record the compatible fallback you expect and check that the channel's release rules can select it. Dashboard rollback does not pin an arbitrary replacement. Verify the recovery procedure on representative release builds. Pausing future selection does not immediately remove code already running on a device. Rollback also cannot undo an external side effect or a destructive data change by itself. Include a separate recovery plan for those changes.

The [rollout and rollback guide](/resources/staged-rollouts-targeting-rollback) explains selection, fallthrough, and local startup protection. Link that operating model rather than asking the approval note to reproduce it.

## Treat uncertain results as a reason to inspect

After an approved action, capture the observed server state and compare it with the intended state. Then monitor the candidate under the agreed plan. A successful tool response confirms its own operation, not application health or fleet-wide completion.

If an upload reports an artifact-copy failure after the server accepted it, inspect the server before retrying. Bundle Drop's [artifact result documentation](/docs/ci-cd#artifacts-and-the-result-file) calls out this distinction. Repeating an upload can create another release instead of repairing the missing local evidence.

Apply the same discipline to a release action that times out: read the current state before deciding whether to retry. Report what is confirmed, what is uncertain, and what remains blocked. Do not convert an ambiguous response into either a success claim or an automatic second mutation.

## Approve the evidence and the next step

A human can make a useful release decision when the record connects a reviewed change to an exact artifact, a compatible tested binary, a defined audience, and a recovery plan. The agent's job is to assemble those facts and make missing evidence visible.

Keep the approval bounded to the next action. A successful initial slice can support a later increase, but it should not silently become permission to widen the release. That boundary lets automation do the preparation while the person approving production can explain what they approved and why.
