“I’m going to unleash a swarm of agents.”
Said nobody. Ever.
(Except for a handful of 22 year old AI influencers on YouTube.)
And to be honest, I’ve been skeptical of AI terms like sub-agents and Ralph Loops.
So when I learned about a little feature in Claude Code called workflows it seemed like a harmless experiment.
Until it unleashed the sub-agents.
Why AI struggles with multi-step tasks
Imagine asking AI a simple question:
Design me 3 investment portfolios and pick the best one.Easy peasy?
Not so fast. A prompt like this leads to a lot of “context contamination.” As AI is performing this task it’s looking at everyone else’s answer.
Not only does Portfolio A know a lot about Portfolio B — but the “judge” has also had a say in the creation process.
Imagine you could “ring-fence” each step of this workflow. You can think of it as 4 separate LLM sessions. One for each portfolio, and a fourth for the judge.
Where each one is completely independent from the prior one.
What if the quality of the ou…
