When I first moved from New York City to Philadelphia, I was starved for social interaction. So naturally, I did what any adult would do.
I downloaded Hinge.
Just kidding. I downloaded Meetup. And for my first social outing, I went to a writers’ meetup. Now the funny thing about a group of adult strangers meeting is that to kill the awkwardness we end up playing childhood games as ice breakers.
And the game we played at this meetup is called “Pass the Story.”
To refresh your memory, about the game, here’s how it works, there’s a sheet of paper. One person writes a line and passes it on. The next person reads what’s written and adds the next line. This continues until, collectively, the group creates a story. Simple, right?
Now here’s the interesting part: If you’ve playe…
When I first moved from New York City to Philadelphia, I was starved for social interaction. So naturally, I did what any adult would do.
I downloaded Hinge.
Just kidding. I downloaded Meetup. And for my first social outing, I went to a writers’ meetup. Now the funny thing about a group of adult strangers meeting is that to kill the awkwardness we end up playing childhood games as ice breakers.
And the game we played at this meetup is called “Pass the Story.”
To refresh your memory, about the game, here’s how it works, there’s a sheet of paper. One person writes a line and passes it on. The next person reads what’s written and adds the next line. This continues until, collectively, the group creates a story. Simple, right?
Now here’s the interesting part: If you’ve played this game before, you already understand ReAct.
But... what is ReAct?
ReAct is a model Observing + Reasoning + Acting, (or you can also say, seeing, thinking, and doing) to solve a problem step by step until it reaches an end goal.
To understand the words “observing, reasoning, and acting” better, let’s map our game “Pass the Story” game to these words.
The first line written on the page?
That’s the query.
When a person receives the paper, they:
- Observe. The person sees what the story so far says.
- Reason. The person thinks about how to continue.
- Action. The person writes the next line and pass it along.
- The person passes the paper to the net person till the reach their goal of having a completed story.
Now here’s the most important part.
The person you pass the paper to has no idea what the story is until they read it.
And just like the players in the game, models are stateless. They have no inherent memory. They don’t remember your previous API call. They don’t remember your previous conversation.
They start fresh every single time.
Like Ghajini. (Yes, the Bollywood movie.)
Passing that sheet of paper forward, the full story so far, is exactly like sending the conversation history back to the LLM in your next API call.
The story so far in the paper is the context.
Now, ReAct is awesome. Most agentic AI systems lean on it. Heavily.
But it has it’s goddam flaws.
1. The "Infinite Loop" Trap.
Imagine you’re playing this game with a toddler who only knows one sentence:
_“Cats and dogs are animals. Animals are cats and dogs.” _
And they keep writing it. Over and over again. Not cause they mean to. They just have limited vocabulary.
In AI, this happens too. It’s called infinite looping.
Sometimes the observation doesn’t provide new useful information. The agent reasons the same thing. Takes the same action. Gets the same result. Repeat.
Solution? You set a limit.
For example: “Only allow 5 iterations.”
2. Context Window & "Token Fatigue"
Imagine, people are really enjoying the game and the story becomes 100 pages long.
The next person can’t possibly read all of it before adding their line.
They’ll skim. They’ll miss context. They’ll get tired. LLMs have the same limitation, and it’s called context windows.
Every Observe → Reason → Act loop adds more tokens to the prompt. This:
- Increases cost
- Slows response time
- Risks hitting context limits
The solution? Context pruning
Instead of passing the entire history every time:
- Keep the last few steps
- Maintain a high-level summary of earlier steps
3.ReAct vs. ReWOO (Plan First, Execute Later)
“Pass the Story” is fun. But you know what’s more efficient? Everyone agreeing on an outline before writing.
In ReAct, you think after every step.
In ReWOO,you think deeply once, generate a structured plan, and then execute it.
It’s the difference between:
- Improvising a story line by line vs.
- Agreeing on the plot first, then writing chapters.
Both are useful. One is reactive. The other is strategic.
Opinion: THINGS ARE MOVING FAST! I feel like I am on a never ending roller coaster ride!
And, I believe that ReWOO is what the IBM video meant by "2026 is the age for multiple agents". Maybe towards the end of this year, this blog won’t even be relevant. :’)
But in the mean time, I hope it was a decent insight into how the ReAct loop works and that you won’t be shy to reach out about your thoughts!
xoxo!
Rolling credits:
- The awesome people of Philly write up group who have welcomed me!
- LLMs