For the library I'm working on I finally getting the power of running multiple agents in parallel. The problem I discovered is that the ping that the agent sends when the job is done is affecting my attention span, like chat app or email pings.
A while ago Boris Cherny, who created Claude Code, mentioned on X that he had 8 tabs with agents open.
The question I have is how many of those tabs are forgotten for days or even longer?
I found out 4 agents running in parallel is my maximum, while maintaining a workflow that is not doing work between job-done-pings.
More than 4 agents gave me more pressure to check when they are done. Maybe over time I will develop a higher tolerance, but for now that is my tipping point.
I can understand that people that are more into AI will start longer running tasks. But I'm running short tasks, because those get the best results.
Before AI came onto the scene, many developers had the idea two monitors were better than one. But after a while a lot discovered the bigger screen size is just a way to add more distractions.
Two screens are good when text needs to be instantly checked against a visual representation. But for most developers that instant checking ability is not needed. We have tests that do the checking for us.
When we were forced to stay at home. One of the positive conclusions was that for mentally intensive work we needed a time were nothing distracted us. And now people at the office also set blocks of do-not-disturb-time in their calendar when they need that distraction free time.
And here we are introducing a new attention seeker with multiple AI agents.
While running AI agents in parallel will make you more productive. This also can make you lose oversight when you don't control the workflow.
How do you handle multiple agents at once in your workflow?
Do you feel like a person that can multitask? Should we multitask?
Top comments (2)
Why not use 2x speed from codex and run one agent at a time, it's soo fast in writing code that this can't be the bottleneck anymore.
Does It run faster because of the LLM or is the agent loop faster or both?
At the moment I'm selectively webscraping, so even when the code generation is fast, retrieving the information could be a bottleneck.