Run agents in parallel, decide one at a time

Several agents can build at once. Your attention cannot be split the same way, so bring each piece of work to one clear decision.

At Blume, we run many agents at once and we do not watch them. The limit is not how many agents you can start. It is how many decisions you can make well.

What goes wrong without it

Four terminals

You start four agents and keep four terminals open. One stops after six minutes to ask whether it may run a command, and you do not notice for half an hour.

Two finish at once. Two edit the same settings file. By the end of the day you have switched context thirty times and reviewed two changes in a hurry.

Four deliveries

The same four changes are four deliveries. Each agent works through its plan and checks without you.

One list shows what waits for you: one question, one outcome review. You answer the question in a minute, open the review, see the evidence and approve.

What to do

  • Let agents run the loop. An agent comes to you with a result and its evidence, not progress reports.
  • Make every request a decision: one question, with the options, a recommendation and the context.
  • Do not block on what can be assumed. See Prefer a note over a block.
  • Take decisions one at a time. Start with whatever unblocks the most work.
  • Separate the work. Each agent gets its own outcome and, where they share code, its own branch or working copy.
  • Stop adding agents when your queue grows. The fix is smaller deliveries or better evidence.

In Blume

  • The deliveries list and board show each delivery’s phase and remaining tasks.
  • Requests gathers the questions and approvals agents asked of you. The sidecar keeps this beside your work.
  • Agents shows each conversation and its deliveries. Agents in a project can message each other.