Running several coding agents at once fails on attention, not on capacity. Six sessions you cannot review are worth less than three you can. This routine keeps parallel work reviewable by one person.
Before you start
- Two or more coding sessions running on the same machine
- Agent Island visible in the menu bar or system tray
- A merge order you are willing to enforce
1. Partition the work before you parallelise
Give each session a scope that does not overlap the others — ideally a distinct working directory. Two agents editing the same files produce conflicts that cost more than the parallelism saved, and neither agent knows the other exists.
2. Let state priority decide where to look
Sessions waiting for you outrank sessions still working. The island surfaces the ones that need attention rather than the ones that changed most recently, so a glance answers "who is next" without opening anything.
3. Choose your two island slots deliberately
The compact island carries up to two provider slots. Pick the two you are actively supervising in Settings → Providers; the rest keep full usage rows without competing for the glance.
4. Require a handoff packet
Ask each agent to end its turn with three things: what changed, what was deliberately not done, and what it is unsure about. The alert tells you a turn ended; the packet tells you what to check. Without it you re-read the whole diff to find the interesting part.
5. Review by risk, not by arrival order
Take the highest-risk change first even when another finished earlier. Arrival order is a queue artefact — it reflects task length, not importance. Migrations, auth, and anything touching data go before cosmetic work regardless of who finished first.
6. Make dependencies explicit before merging
Parallel agents cannot know about each other, so ordering has to come from you. State which session depends on which before the first merge, then merge in that order and re-run the affected checks.
7. Tune concurrency to your review rate
If alerts arrive faster than you can review, the answer is fewer sessions rather than muted alerts. Concurrency should be set by how fast you can review, not by how many terminals fit on screen.
If it does not work
Reduce concurrency. Muting alerts converts a visible backlog into an invisible one.
The partition in step 1 was wrong. Resolve manually, re-scope, then restart — do not let both continue.
Every row and alert carries a session label and working directory. Name directories meaningfully and the labels do the rest.
That is a missing handoff packet, not a monitoring gap. Add the requirement to the prompt.
Turn off Show thread details in Settings → Status guide; the alert still fires without naming the project.
For the priority rules behind step 2, read multi-session state priority.
← All posts