<!-- Keep humans in control of conflict resolution — Merget · https://merget.ai/blog/keeping-humans-in-the-merge -->

# Keep humans in control of automatic conflict resolution

Start in Advisory, inspect the evidence, then choose how far Merget can act. A practical guide to modes, merge requirements and a first repository.

Merget team · Nov 13, 2026 · 5 min read

Automatic conflict resolution should give a team more control over integration work. Merget starts in [Advisory mode](https://merget.ai/how-it-works#modes-overview) so you can inspect conflicts before allowing it to change source code or merge pull requests. From there, Queue and Autonomous modes let you choose how much of the merge process should run on its own.

The distinction matters for teams using AI coding agents. An agent can propose a good local change without seeing every other open pull request. The semantic merge queue sees where those changes meet, resolves the conflicts in order and presents the result for your configured checks and reviews. People still choose the repository, the mode and the requirements for landing.

## Start with Advisory and inspect the evidence

A newly connected repository starts disabled and in Advisory mode. Once enabled, Advisory reports findings without changing source code or merging pull requests. It makes no model requests. This gives the team a way to examine the PR graph, the conflicting code and the explanation before delegating resolution.

Use known examples for the first evaluation. A renamed helper with a new caller, a removed definition still referenced by another branch, and two genuinely independent pull requests make a useful small set. Ask whether the report identifies the right changes and whether the evidence explains the collision. Count useful findings and missed cases, not just the total number of alerts.

## Queue mode adds resolution while you approve the merge

In Queue mode, Merget can resolve conflicts and prepare work against the expected results of the pull requests ahead of it. A person still approves the merge. That separates two decisions: whether the queue may propose and publish a fix, and whether the combined work should reach the target branch.

A team can inspect the proposed result, run its usual checks and review whether the fix preserves both changes. For a rename, did the new caller move to the correct API? For a policy change, does the combined behavior match the intended outcome? The goal is to replace repetitive rebase work with a concrete result that can be assessed.

## Autonomous follows the requirements you configure

Autonomous mode can land work through the configured merge requirements. Merget checks the actual target before landing because the branch may have moved since preparation began. If that movement invalidates the prepared result, the work needs to be checked or prepared again.

Automation does not remove the need to choose appropriate tests or review rules. It makes the agreed process repeatable. Inspect repository settings before enabling it, including any explicit override controls. Authorized users can make separate decisions through those controls; the ordinary resolution path should not be confused with a guarantee that every possible setting forbids every override.

## Understand what resolution can read

Queue and Autonomous modes can use an agent when a fix needs judgment. The repair agent works in a complete checkout of the repository being prepared, so it can read any file there and run commands. Code it reads and the conflict context may be sent to a model service.

That is a reason to settle data requirements before enabling resolution. Advisory’s no-model behavior is a useful starting point; agent-assisted repair is a separate decision for each repository. Choose the repositories the GitHub App can access deliberately, and read [how Merget handles data](https://merget.ai/security).

## A first repository needs a clear success criterion

Start with one repository where parallel agent work regularly collides. Choose an integration owner who can compare Merget’s findings with the code and explain the intended combined behavior. Bring representative pull requests, CI configuration, review rules and any model-use restrictions. Agree which mode to use before you begin.

Useful success criteria describe outcomes: fewer repeated rebases, less engineer attention spent reconstructing conflicts, or more changes landing with the expected behavior. Raw branch counts alone may not describe equivalent work, and a faster run is not useful if it omits the feature the team needed. Compare the content integrated and the checks passed alongside time and cost.

Include a difficult case that should require a person. The team should be able to recognize an unresolved or ambiguous fix and know what to do next. A tool that clearly hands off a case it cannot resolve is easier to operate than one that makes every outcome look successful. That boundary is part of evaluating the workflow.

## Keep a practical stop and recovery path

Owners can stop new automatic merges from starting. A merge already in flight may still finish, so pausing is not an instant rollback. Disabling or removing the GitHub App’s access stops Merget taking part in the affected repositories. Discuss who owns those controls and how the team handles an unexpected result before the queue becomes routine.

When a result is wrong, inspect the prepared change and the finding behind it, then correct the code through the team’s normal workflow. Preserve a small example for improving the evaluation. Whether the fix came from a person, a deterministic resolver or an agent, the combined program and its intended behavior remain the basis for the decision.

Move between modes when the evidence supports the next step. Advisory establishes whether the findings help. Queue establishes whether the resolutions preserve the work and are easy to review. Autonomous applies the merge process you have chosen. The purpose of the progression is straightforward: spend less time repairing the same integration problems while staying deliberate about what lands.

[See how Merget resolves conflicts in queue order](https://merget.ai/how-it-works), or [sign up free](https://sema.merget.ai/login) and connect one repository.

### How a semantic merge queue works

How a semantic merge queue follows code relationships across pull requests, resolves conflicts in queue order and checks the combined result before landing.

Merget team · Nov 13, 2026 · 5 min read

[Open](https://merget.ai/blog/what-makes-a-merge-queue-semantic)

### Why a clean merge can still break main

Two pull requests can pass alone and merge without text conflicts while breaking the combined build. Here is a small example and the workflow that catches it.

Merget team · Nov 13, 2026 · 5 min read

[Open](https://merget.ai/blog/why-a-clean-merge-can-still-break)

### Eight ways branches collide

Eight concrete merge-conflict patterns, from renamed helpers to competing writes, and how to decide which need a fix or human judgment.

Merget team · Nov 13, 2026 · 5 min read

[Open](https://merget.ai/blog/eight-ways-branches-collide)

[All posts](https://merget.ai/blog)
