How to Roll Out Switchboard Gradually Without Automating Everything

Do Not Start With The Whole Backlog

Switchboard is most useful when it turns clear Linear issues into reviewable pull requests. That does not mean the first rollout should watch every project, every label, and every repo.

A gradual rollout is safer and faster to learn from. Start with one lane, prove the issue shape, tune the route, review the pull requests, then expand.

The goal is not to automate everything. The goal is to automate the work that is ready.

Phase 1: Manual Runs First

Before Switchboard, run a few Claude Code or Codex sessions manually through Junction. Pick the kind of work you eventually want to automate.

Manual runs teach you:

• Which prompts work.
• Which repo instructions are missing.
• Which tests the agent should run.
• Which files the agent tends to touch.
• Which review notes humans need.
• Which machine should run the work.

If manual runs are chaotic, automation will be chaotic faster.

Phase 2: One Repo, One Issue Type

Choose one repository and one issue type. For example:

Repo: packages/site
Issue type: docs and copy fixes
Linear signal: switchboard-ready label
Daemon: team-vps
Concurrency: 1

This narrow lane gives you clean feedback. If a run fails, you can inspect the issue, route, daemon, and repo without a giant matrix of possibilities.

Phase 3: Add A Ready Gate

Use a dedicated Linear status or label to mark issues ready for automation. Do not let Switchboard pick up raw backlog items.

An issue is ready when it has:

• Target repo.
• Clear goal.
• Acceptance criteria.
• Test or validation notes.
• Constraints.
• Review owner.

The ready gate is the most important rollout control. It keeps automation from becoming a cleanup job for vague tickets.

Phase 4: Review Every Output

In the first rollout, review every output carefully. Look for patterns:

• Did the agent stay in scope?
• Did it run the expected checks?
• Did it open a useful PR?
• Did reviewers understand the summary?
• Did the issue contain enough context?
• Did route instructions help?

This is not distrust. It is calibration.

Phase 5: Expand One Dimension At A Time

When the first lane is stable, expand only one dimension:

• Add one more issue type.
• Or add one more repository.
• Or add one more team.
• Or raise concurrency slightly.

Do not add everything at once. If quality drops, you need to know why.

Keep a short change log for the rollout itself. Note when you added a route, changed concurrency, edited instructions, or expanded the eligible issue set. When a later run fails, that history helps you separate a bad issue from a bad rollout change.

A Good Rollout Sequence

Week 1: Manual runs for the target issue type
Week 2: Switchboard on one repo with concurrency 1
Week 3: Tune issue template and route instructions
Week 4: Add one adjacent issue type
Week 5: Add another repo only if review quality stays high

The exact timeline can vary. The sequencing matters more than the dates.

What To Keep Manual

Even after Switchboard is working, keep some work manual:

• Ambiguous product changes.
• Security-sensitive edits.
• Large refactors.
• Cross-repo architecture work.
• Migrations.
• Work without a clear reviewer.

Manual does not mean anti-agent. It means a human should steer the session directly before it becomes a pull request.

Tradeoffs

Gradual rollout can feel conservative. You may have more automation capacity than you use at first. That is fine. The fastest way to lose trust in agent automation is to flood reviewers with noisy PRs.

The other risk is never expanding. Set a review rhythm. If the lane is healthy, add one more useful slice.

The best signal is reviewer behavior. If reviewers trust the PRs enough to review them normally, the lane can grow. If reviewers start ignoring them, shrink the lane before adding more automation.

Where Junction Fits

Junction gives you the manual control surface first: live output, approvals, diffs, Git state, and notifications. Switchboard adds the automation layer: Linear pickup, route settings, isolated worktrees, and issue-to-pull-request flow.

If you are deciding when to automate, read Manual AI Agent Runs vs Switchboard Automation. If you are configuring the route, read Configure Switchboard Route Settings for Team and Repo Routing. Compare plan fit on the pricing page.