More people seem to run two or three coding agents in parallel now: one on a bug, one on tests, one on a refactor.
How do you keep an eye on them? Separate terminals, IDE panes, a second monitor, notifications?
And what is the most annoying part when two of them are waiting on an answer from you at the same time?
Top comments (6)
A status line in each terminal. With 11 sessions open, 6 showed the same line, because sessions in worktrees only displayed the parent repo's name. The second line now carries the session's tag and the worktree name. The branch cache was a separate bug: one global file for every session, so a terminal could show another repo's branch. It's keyed by directory now.
I find the hardest part isn't monitoring what each agent is doing; it's managing the coordination overhead when they all need attention at once.
Separate terminal panes work well visually, but I'd give each agent a clearly isolated task and workspace. For example, one handles a bug, another writes tests against the expected behavior, and a third investigates a refactor without touching the same files. Otherwise, parallel execution can quickly turn into merge conflicts and agents overwriting each other's assumptions.
I'd also want every agent to expose three things at a glance: current task, last meaningful action, and whether it's blocked or waiting for approval. A stream of token output isn't particularly useful when you're supervising several agents.
The biggest improvement is probably asynchronous handoffs: agents should preserve their findings and specific questions so you can review and respond without reconstructing the entire conversation. Parallel agents save time only when the coordination cost stays lower than the work they take off your plate.
Isolated task plus isolated workspace is the right split. The overhead I keep running into is when two of them need a decision in the same minute. Do you queue those, or answer whichever came first?
The failure mode I'd watch for with one agent on a bug, one on tests, and one on a refactor is them editing the same files in the same checkout — the refactor agent can quietly undo what the bug agent just fixed. I'd give each one its own git worktree or branch, and when two of them are waiting on you at once, check the diff of the one you're about to answer before you reply, not its summary. That way a waiting agent can't get a yes based on work another agent has already overwritten.
two agents, one decider. the watching is easy, the deciding is the bottleneck.
"The watching is easy, the deciding is the bottleneck" sums it up. Thanks for this.