To kick off the month of August, we’re covering two of the hottest topics in AI right now.

Loop and Graph Engineering.

Over the past few months, we’ve shifted away from prompt engineering, and we’ve entered an environment of creating “loops.”

Aka creating systems where agents complete tasks end-to-end on your behalf with little manual prompting intervention needed.

Master the principles discussed in this publication, and you’ll be on your way to automating your life with Claude in just a few minutes.

This is everything you need to know about loop and graph engineering explained simply.

Loop Engineering Explained In 30 Seconds

Up until this point, working with AI has looked something like this:

You send a prompt → AI responds → You iterate → Repeat

This is classic prompt engineering, and what loop engineering does is automate this entire cycle.

Instead of typing message after message, you set the goal once, and Claude
keeps working on it by itself.

Claude does the task, checks its own work, and only comes back to you when it's finished or stuck.

Old way: you prompt, Claude answers, you prompt again, forever.

New way: you describe the goal once, and a loop works, checks itself, and reports
back.

Running Your First Loop

Getting started with Claude loops is simple.

Open Claude Code or the desktop app and type: /loop

Then describe what you want the loop to do.

Example: “/loop research the robotics sector and find me a few potential
investment opportunities and some basic exposure.”

Scheduled loops

The cool thing about Claude is you can actually run loops on a schedule.

Example: “/loop every morning at 9 am, summarize the overnight AI news.”

There are two things to get right before you deploy any loop:

Claude loops work best with a clear end goal deliverable.

Things like: research reports, analysis reports, and so on. Always specify what format and end goal you want the loop to complete.

You also need to tell Claude when to give up so it doesn’t run all night or burn through token costs.

Something like: "If this fails five times, stop and tell me."

Where Loops Hit A Wall

The problem with loops is that once the work gets complex, they very easily hit a wall.

In a loop, all actions happen one step at a time. This is often described as a “line” of work.

Loops can't run multiple tasks in parallel, and this is the biggest limitation.

For example: if your task has four independent pieces of research, a loop does them one after another.

This is where graph engineering comes in, and it essentially allows multiple loops to run concurrently.

Graph Engineering Explained In 30 Seconds

A graph is your workflow built as a “network” instead of a line.

Instead of one agent grinding through steps in order, you set up a structure, and each node has a specific job.

The nodes connect, work splits off to run in parallel, then things converge back together.

Here's the shape we recommend everyone start with:

Running Your First Graph

You don’t need to be technical for this. You just need to follow four simple steps:

Step 1: Audit your loops

Look at what you already run. Which workflows take the most steps, retry the most, or cost the most to finish? You're hunting for one or two bottlenecks worth fixing.

Step 2: Find what can run in parallel

Map out which tasks in your loop can run in parallel.

A good rule of thumb is if a task isn’t reliant on another task, it can usually run in parallel as it doesn’t need anything to get started.

Step 3: Draw your graph

This is a pro tip we’ve found helpful:

To deploy a graph, we recommend actually drawing it out.

This could be hand-drawing it, or by connecting the Excalidraw MCP and having Claude draw it for you.

Example: A graph drawing built for researching 5 competitors

Step 4: Hand Claude the graph

Screenshot your drawing, paste it into Claude, and tell it to deploy a graph:

A graph is what you get when you give subagents defined roles, run them in parallel, and write a rule for where failures go. It’s that simple.

Final Tips

Some final tips for both loop and graph engineering

UnslothAI recently revealed versions of Kimi K3 that can run entirely on private hardware.

→ Download the Kimi K3 GGUF files through Hugging Face

→ Run them with Unsloth Studio or llama.cpp

This is a massive model and would likely cost $20,000+/mo to run locally.

Nonetheless, UnslothAI has a cool guide worth reading about how local models like Kimi K3 actually work:

Unsloth’s local guide

Alibaba just released Qwen3.8-Max, its strongest model yet for coding and professional work.

→ 1M-token context window

→ Multimodal support

→ Open weights coming next week

Try it now in Qwen Chat. https://chat.qwen.ai/

GPT-Live now powers ChatGPT Voice.

It can listen while it speaks, so you can interrupt, pause, or continue talking without waiting for separate turns.

Open ChatGPT, tap the Voice icon, and try having a natural back-and-forth conversation.

xAI has released Grok Voice Think Fast 2.0, its new model for live voice agents.

→ Responds in ~0.7 seconds

→ Reasons while speaking

→ More accurate in noisy environments

→ Improved transcription across 24 languages

You can try it free in xAI’s Voice Agent Builder, or access it through the API.

https://x.ai/voice

OpenAI is teasing its next major model after an internal version produced ten advances in mathematics and theoretical computer science.

Some people are already calling it GPT-6, but OpenAI has not confirmed the final name or release date.

AI 2027 is a paper detailing a month-by-month prediction of AI progression.

What it predicts comes next:

→ AI labs deploy thousands of agents to improve their models (agents become mainstream)

→ AI progress will exponentially compound from here as AI helps build better AI (exactly what we saw with Fable 5 “training” Kimi K3)

Read the full research here, and track which milestones happen and which do not.

Our barbell strategy for cutting token costs by 80%:

Rumors now point to Grok 4.6 dropping in the next two weeks.

xAI has not published final benchmarks or confirmed the exact release date, so this remains a watch item.

We put together seven AI books worth reading.

Covering AI automation, AI bottlenecks, philosophy, and more.

If you made it this far, thank you for reading, and we hope you found this week’s edition valuable.

If you enjoy reading, please forward our newsletter to someone you think would benefit from it.💙

Our promise to you: Every Wednesday, at 7 am EST, we’ll cut through the AI noise and send you human-curated AI content to make sure you stay ahead.

YouTube

Live Now: How I turned ChatGPT Into My Real-Time Voice Assistant (by Miles)

X (Twitter)

Live Now: Context Engineering for Claude 5 Models

Free AI Asset Library

For all the AI prompts, cheat sheets, and PDF guides mentioned on the AI Edge YouTube channel, grab them by browsing our free assets library here:

Interested in sponsorship or a partnership? Get in touch at [email protected]