Background Tasks: Running Multiple Agents In Parallel

A background task is work the agent runs without blocking your interactive session. You kick off a long search, a multi-step refactor, or a test run, and keep working on something else. When it finishes, you get a notification with the summary.

When background tasks earn their place

1. 1Long codebase searches that would otherwise block the session for minutes
2. 2Test suites that take 5+ minutes to run
3. 3CI-style validations (lint + test + typecheck) on a series of changes
4. 4Dependency upgrades where you want progress reported as it goes
5. 5Parallel exploration: try three approaches to the same problem in parallel

Where it goes wrong

• Spawning 5 background tasks and forgetting which is which
• Tasks that need confirmation halfway through and just hang waiting
• Tasks that touch the same files as your interactive session — race conditions in your own diffs
• Background tasks consuming your token budget while you sleep on it
• Workflows that conceptually need to be sequential but get parallelized for the wrong reasons

A two-track workflow

Apply: queue one task tonight

1. 1Pick a search-heavy task you've been putting off (audit all Bash commands in CI scripts, list every task list in src/)
2. 2Spawn it as a background task with a clearly defined output
3. 3Switch to your active work; ignore until notified
4. 4Read the result; decide what to do with it

The big idea: background tasks add throughput when used for clearly separable, recoverable work. Don't background ops; don't poll.