Loops are In, Prompting is Out

[ben]

According to Boris Cherny, the father of Claude code, he is no longer prompting, but writing loops: “I don’t write prompts anymore, I have loops running that prompt Claude… My job is to write loops. I uninstalled my IDE, I wasn’t using it.” - Boris. Full interview here.

Peter Steinberger, the creator of OpenClaw, said something similar a few days ago:

So, what does that mean for people building software now?

These types of “loops” are very interesting. We’ve been spammed about Agent Loops and for loops for a while, but now we’re getting into loops that run on a specific interval with a particular objective in mind (I might call this a workflow that runs on an interval).

But, Ben, why would I run a loop?

Well, glad you asked! It takes a bit of a mind switch and it clicked for me when thinking about the giant list of tasks swallowed away in Jira’s backlog.

Going through and picking out which bugs to tackle, then spinning up a session for each one takes mental overhead in order to accomplish, especially when the agent would ask “should I create a PR for this” or “how should I commit this?” every single time.

But with a loop, you’re creating a structure for it to follow for every single item.

This reduces the mental load almost entirely. The work is done with a process that you describe, and you can designate specific handoffs, instead of having to decide how to manage it every single time.

And it’s freaking awesome so far.

I’d be curious to see how other teams might start re-thinking their work… every meeting, note taken, Figma comment/update, bug captured, feature requested, you can write a loop for that to create a structured process that leaves a designated handoff for you to review.

[Bryan]

Good stuff Ben. When everyone has access to roughly the same models, the idea behind the loop I think is to get compound advantages from AI’s strengths (while also reducing mental burden)

Teams that get an advantage are going to be focused on:

• organizational memory
• decision systems
• operational workflows
• shared context
• structured processes
• how work moves through the company

Anthropic seems to be mainly interested in:

1. agent workflows
2. organizational context
3. decision loops
4. feedback systems

Human → System → Agents → Feedback → Decisions → Agents

The value here is the memory system.

It has me thinking about Glare in this context… how do we get to the organizational memory?

I think when we first started Glare, we were focused on the information and shared understanding (of structured user data)…but realized that we needed a way to operationalize the decisions. Skills led to agent operations recently, which is now leading us to memory.

Information
↓
Shared Understanding
↓
Decision Systems
↓
Agent Operations
↓
Organizational Memory