Codex, Jules, and the Dawn of Async Agents

Codex and Jules are the first AI products I've felt I could trust to go off and accomplish a task async. You should try them. They'll increase your faith in what LLMs can do in one-shot.

Until now, chat-based AI interactions have felt like single-threaded processes. You ask a question, wait for a response, and proceed linearly through one task at a time. 

Async agents like Codex and Jules, on the other hand, let us treat AI as a true second brain. You fire off a task, forget about it, let it run in the background, and return to find results.

This feels like expanding your cognitive bandwidth. Ask Codex to refactor code during a walk, then check it when you return to your computer. Ask Jules to explore implementation for a feature idea that crosses your mind during deep work, then review it later without losing your train of thought.

From Codex and Jules, we can extract some general principles for building effective async agents:

• Plan then execute: Effective async agents show you their plan before starting work. When you give Jules a task, it presents an execution plan you can review and iterate on before it begins execution. This prevents the anxiety of wondering whether the agent understood your request correctly.
• Structured review: Async agents work best when they plug into workflows you already trust. In software development, the pull request model provides a natural pattern — agents contribute work that gets reviewed before integration. An async agent is only as reliable as your review process.

As async agents mature, I'm most excited about developments that will make working with them feel even more natural:

• Intelligent Checkpointing: Future agents will recognize when human input would be valuable and proactively pause for guidance.
• Multi-Branch Exploration: Agents will evolve to explore multiple approaches simultaneously, presenting you with options rather than committing to a single path. Imagine getting back: "Approach A worked (70% confident), Approach B needs your input at step 3, Approach C requires different tools."
• Agent Inbox: We'll see various ways to track what agents are working on, what's finished, and what needs attention. This becomes especially powerful with multi-branch exploration—managing multiple parallel efforts without losing track of discoveries.

One thing you notice when using async agents is that reviewing is much easier than generating. This generator-reviewer asymmetry means that even imperfect async agents can provide a ton of value by handling initial generation tasks. You really should try firing off more tasks to Jules / Codex.