Welcome back. Every AI lab (big or small) is working on getting the AI coding harness right. It's a win for developers either way. A small lab in Silicon Valley just outperformed Claude Code and Codex on some of the toughest benchmarks, and builders are piling on fast.

Also: Meta’s new coding agent, token wastage calculator for coding agents, and how to run DeepSeek V4 in Codex.

Meta ships a coding agent for large codebases: The social media giant just dropped Muse Code in beta, a terminal coding agent powered by its newest model, Muse Spark 1.2. It plans changes, writes code, and validates results across large repos. For massive tasks, it automatically splits the work across persistent sub-agents running in parallel. Developers can also access the model directly via the Meta Model API, which just expanded its global footprint. Check out what you can build with Muse Code.

Prime Intellect’s new coding harness outperforms Claude Code and Codex: The Silicon Valley based AI lab just unveiled Prime Agent, a self-improving harness designed for long-running autonomous coding tasks. It allows for indefinite sessions without context rot by spinning out parallel sub-agents and converting repeated fixes into reusable skills. The lab claims Prime Agent outperforms existing harnesses like Claude Code and Codex on long-range benchmarks, even hitting a 95.5% ARC-AGI-3 score to beat the human-expert baseline. 

Figure AI’s computer use agent claims to rival GPT 5.4 and Opus 4.8: The robotics company just unveiled Handoff, an agent designed to handle web tasks end-to-end. By clicking and typing through live sites, it can navigate platforms like Walmart and LinkedIn that don't have public APIs. For developers, that could open a new platform for building agentic apps on services that were never built to be automated. Watch it in action.

Training data for physical AI is hard to get right: every new task needs exact visual examples, diverse environments, and edge cases.

At Machina in July, Bright Data’s CEO Or Lenchner showed how robots could query searchable video data by prompt and timestamp.

Bright Data delivers targeted video feeds for VLA teams pre-cut MP4 clips with precise timeframes and structured metadata, ready for training pipelines.

Get Sample Clips

Last week, AI chipmaker Cerebras shared a breakdown of how they built their internal knowledge base. It quickly racked up 2.9 million views on X. The system already handles over 15,000 questions a day, despite launching only three months ago. 

The data lives everywhere. Company knowledge is scattered, and every quarter someone proposes the same fix: record everything in one platform. That single source of truth rarely works: people create information wherever it's easiest. Design debates live in docs, engineering talk in Slack, code in GitHub, and status in Jira. Nobody wants to discuss a pull request in Google Docs.

Meet the data where it lives. Their strategy was unique. Instead of asking employees to change their methods, they decided to pull data from each of these platforms. The system does three things: collects internal data, makes it searchable, and controls who sees what. Every source, from Slack threads to chip design files, lands in one Postgres table with a single search interface.

Slack messages were the hard parts. You can find raw text by keyword immediately. But keyword search misses meaning, because two engineers can describe one problem in different words. So an LLM rewrites each thread into a clean record: the question, a summary, the fix, and the systems it names. The team saves and searches that clean record, and the raw thread can still be searched using keywords.

Code had the same problem. A coding agent like Claude Code can grep a repository, but grep finds only known words. Research from Cursor showed that search by meaning finds the code that keyword search misses. So Cerebras embedded its repositories too and re-embedded only the changed code on each commit. Now every source shares one table, and a lightweight LLM picks the right sources for each query.

P.S. To begin setting up a knowledge base for your engineering team, read this post from Cerebras, and if you want to build one for personal use, start with this tutorial.

How to build web apps for Meta Ray-Ban Displays: You’ll learn how to build lightweight web apps for Meta Ray-Ban smart glasses. This tutorial covers everything from designing for a 600x600 resolution and implementing gesture controls to managing pairing codes and API security.

Token wastage calculator: This tool helps engineering teams cut down on unnecessary AI costs by tracking exactly where tokens are being wasted. By giving you a clear look at how coding agents like Codex or Claude Code use data, it helps you sharpen your prompts and keep your context windows lean, so you only pay for the results you actually need.

How to run a workspace for humans and agents: Self-hosting a Buzz workspace, a Slack-like platform for humans and agents, often involves jumping through a lot of hoops with complex infrastructure. This guide shows you how to easily deploy and run your own private relay using a simple Rust binary and Docker. 

It’s tough to keep track of everything in a large repo, and coding agents often swap out modules without realizing which parts of the code actually depend on them or which tests are supposed to cover them. This engineer shared a Codex prompt that solves this exact problem. 

The thread also includes an AGENTS.md snippet that makes Codex check the map before every code change.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Apple sued OpenAI for stealing hardware secrets. Then OpenAI published private messages that flip the script on who was talking to the accused engineer. Check out the full breakdown.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ 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