When I Reach for Background Agents vs Interactive Ones

The agent landscape has bifurcated. Interactive agents like Claude Code work turn by turn with a human in the loop. Background agents like Codex run asynchronously, often for minutes, and report back when done. Both have their place. Mixing them up costs time.

The Nature of the Work

Interactive agents are best when I need to steer. Exploration, debugging, learning a new codebase, making judgment calls on architecture. I read each response, decide the next move, and we converge together. The loop is tight.

Background agents are best when the work is well specified and long running. Apply this pattern across 20 files. Run the test suite and fix failures. Upgrade the linter config and resolve the new warnings. Tasks where the goal is clear and the path is mechanical.

The Cost of Mismatch

When I give a background agent an exploratory task, it comes back with a big confident PR that is 60 percent right. Reviewing and fixing that PR takes longer than doing the exploration interactively would have. The agent's speed does not help me when I did not know the destination.

When I give an interactive agent a mechanical task, I burn my attention on something I could have delegated. I am the bottleneck in that session and I could be thinking about something more valuable.

Rough decision tree:
 
Do I know exactly what success looks like?
  No  -> interactive (Claude Code, Cursor)
  Yes -> is the work mechanical?
         Yes -> background (Codex-style async agent)
         No  -> still interactive (judgment required)

The Workflow I Run Now

Mornings are interactive. I spend them on design, architecture, and new code with an agent at my elbow. Afternoons I hand off background tasks and work on something else while they run. Evenings I review the background results.

This rhythm doubles my effective output because I am not context-switching between deep thinking and mechanical work. The agents handle the latter while I focus on the former.

The Meta Skill

Recognizing which mode a task needs is itself a skill. It took me about two months of daily use to develop the intuition. Now it is unconscious. Most of what I would call being good with AI agents is really being good at matching tasks to agent styles.

For background agent capabilities, see the Codex documentation. The mode choice matters more than which vendor you pick.