I spent some time this weekend building a choose-your-own-adventure app for my kids. One of those "quick Saturday afternoon ideas" that morphed into me outlining an agentic narration system with sub-agents checking each other's work.

The idea started simple — make a fun, easy-to-play D&D-style game, to keep the kids entertained. Basically, a lite dungeon master powered by an LLM. Pretty quickly I realized that from a single prompt, the LLM was skipping points where it should call for user input and making up those actions on its own.

It worked well enough from a single prompt, but that got me thinking of how to break down each task that goes into telling the story. A director to oversee the story and make sure it follows along the key beats. A narrator to do the creative writing and an arbiter to check the narrator isn't pre-determining user actions. Then a chronicler to summarize the beats and clean out the fluff from the context window to send back to the director and narrator and repeat.

It's a pattern that was inspired by things like the Claude "dreaming" that came to light in the Claude Code source leak. Which was then adopted into the OpenClaw project's own dreaming system as a way for agents to consolidate memories each night.

The greatest benefit from AI tools isn't one-shotting a solution from a single prompt. It's the ability to throw compute at a problem multiple times to sharpen the output using models with reasoning and thinking capabilities built-in.

I can take a complicated task like tracking narrative story with user input and break it down through a variety of sub-agents. That each have their own rules to follow and context windows. By doing that (in theory at least) the output should be better and use up fewer tokens overall than sending everything back to a single context window.

Whether it's an adventure game for my kids or an AI coding assistant, the principle is the same. Don't ask one prompt to do everything. Split it up into smaller, focused chunks with the context to help it make the best prediction of what should come next.

💧 Article
Claude Code's Entire Source Code Got Leaked via a Sourcemap in npm, Let's Talk About it
kuber.studio

Someone forgot to add *.map to their .npmignore and now we all know what's inside Claude Code. The technical details are fascinating, but the real value is seeing how a production AI coding agent actually works under the hood. Especially to see how the people making the harnesses are guiding their models (re: dreaming). Very cool to see the concepts and ideas under-the-hood. Also a good reminder to check your ignore files.

This was also ages ago in AI standards (2 weeks) and a lot of the leaked features (Buddy, Ultraplan, etc.) have been released now too.

💻 Article
Frontier Agent: AWS DevOps Agent
AWS

Listen, I've done my fair share of poking around in AWS and deleting configs I probably shouldn't. (Always have a backup.) But having a DevOps Agent would make me much more comfortable poking around in all the details of AWS. On-call DevOps seems like an excellent utilization of AI computing firepower. Just, you know, be careful with what it can do on production. And always have a backup plan.

✨ Article
Introducing Muse Spark: Scaling Towards Personal Superintelligence
Meta Superintelligence Labs

Meta is getting back into the model game for the first time in a while with their new Muse Spark model. It's the first release from their reworked Superintelligence Lab and by initial accounts seems to be bringing them up to competition with the Geminis, GPTs, and Claudes. Simon Willison had an interesting breakdown of the tools Muse Spark has available to it.

♠️ Article
Claude Managed Agents: Get to Production 10x Faster
Anthropic

Anthropic is taking aim at another AI enhanced product with their own Managed Agents platform. It handles the infrastructure, sandboxing, credential handling, and orchestration bits to make building agents faster and require less technical know-how. It also comes with a side of vendor lock-in. They did share some case studies from Rakuten and Sentry who are already building agents on the platform.

🕵️‍♂️ Article
Welcome to Agents Week
Cloudflare

Cloudflare kicked off their Agents Week. They're making a big play to be the infrastructure behind agents on the web. Over the upcoming week they'll be releasing new things every day as they look to build the infrastructure for the AI-age. They kicked things off on Monday with secure sandboxes for your agents and notes on building a CLI for all of Cloudflare.

Thanks for reading, Jason