Welcome back. Perplexity just spent months building something in total secrecy, and it’s not what anyone expected. It isn’t a new model or a flashy feature update. They just unveiled 'Personal Computer’, an always-on AI system that hooks directly into your local files and apps, powered by a dedicated Mac mini on their secure servers.

Also: A dev ships entire code with 5 autonomous agents, a 16-part series to master DSA and build systems, and Notion just opened up a brand-new engineering role.

Perplexity turns a Mac mini into a 24/7 AI agent: The AI startup just announced Personal Computer, a dedicated Mac mini designed to run 24/7 that connects your local apps and files to Perplexity's cloud. It acts as a digital proxy, managing tools and tasks from any device. Alongside this, they launched four new APIs: Search, Agent, Sandbox, and Embeddings, which are built on the same infrastructure and are now available for external use.

Replit launches its most capable coding agent: The web-based coding platform shipped Agent 4, enabling teams to design on an infinite canvas while parallel agents handle front-end, back-end, and database work simultaneously. From apps and websites to slide decks, everything now deploys from a single project. See it in action.

Google launches its first multimodal embedding model: The search giant just unveiled Gemini Embedding 2 in public preview, its first natively multimodal embedding model based on the Gemini architecture. By mapping text, images, video, audio, and documents into a single embedding space, it removes the need for separate pipelines when building RAG systems, semantic search, or data clustering tools.

Civilization was built on voice. So why do most human-to-AI interactions take place with a keyboard?

In the first-ever ElevenLabs summit, Co-Founder Mati Staniszewski shared his predictions for the future, and spoiler alert: He believes voice will be the #1 interface with GenAI, rendering keyboards useless.

Watch how voice will revolutionize everything.

Mo agents, mo problems. Francesco Bonacci, a YC founder and engineer, posted on X that he was finishing every day drained "not from the work itself, but from the managing of the work." Harvard Business Review researchers studied the pattern and gave it a name: AI brain fry. It’s the cognitive tax that builds up from overseeing agents all day long.

The oversight is the problem. Nearly 18% of engineers already have it. And companies are quietly making it worse. Meta tied performance metrics to lines of AI-generated code. This forced engineers to run more agents to hit their numbers, leading to the constant monitoring that causes fatigue.

Your best people are the most exposed. The engineers getting hit hardest tend to be the highest AI users. They are the ones moving quickly and managing multiple agents at the same time. By pushing teams toward AI adoption, companies are building a retention problem into the same tech they're betting their future on.

The fix is structural. Simply telling engineers to be more intentional isn't the solution. Teams with an organized approach to AI integration experience significantly less strain. Start by documenting what runs autonomously versus where you need a human checkpoint. Sharing this with the team shifts 'brain fry' from a people problem to a design problem, and design problems have solutions.

Sharing AI coding sessions can be a real hassle. Screen recordings are often too bulky, and raw transcripts are almost impossible to navigate. This repo solves that problem by converting Claude Code session logs into self-contained HTML files. They're easy to email, embed in your documentation, or host wherever you need them.

Claude Code stores its transcripts in “~/.claude/projects/”. Just point the tool at your project to get started:

It features an interactive player with speed controls, collapsible tool calls, and chapter bookmarks, all designed to work seamlessly in any browser.

The generated HTML embeds your entire session, including file paths and tool outputs. While secret redaction is enabled by default, it may not catch everything in your source code, so be sure to review the file before sharing it publicly.

How to use /btw and fork sessions in Claude Code: You’ll learn to optimize Claude Code workflows using the new “/btw” command. This feature lets you ask the AI questions inline without interrupting its progress or cluttering the context window. The tutorial also explains when to fork sessions for complex tasks or rewind to fix mistakes.

Context Hub (by Andrew Ng): Coding agents hallucinate APIs and forget what they learn in a session. Context Hub gives them curated, versioned docs and the ability to get smarter with every task.

Evaluating agent capabilities (by Alibaba): Current AI benchmarks focus on quick, one-off code fixes while ignoring the reality of long-term software maintenance. When applying a new continuous integration benchmark, researchers found that even top-tier AI models struggle to maintain code over time without breaking existing features.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team