Welcome back. Competition between OpenAI and Anthropic is getting fiercer by the day — and it’s great news for devs! This time, OpenAI is slashing GPT-5.6 Luna by a whopping 80%. And that’s not the only good news today: GitHub has a handy new tool for engineering teams working with large codebases.
Today’s Brief
A developer’s guide to Harness Engineering
Become a LangChain certified agent engineer
7 ways to cut your token usage (tutorial)
How to build agent-first apps (cookbook)
TODAY IN PROGRAMMING
Click here to check out the new API plans from OpenAI.
OpenAI slashes API prices to win developer workloads: The ChatGPT maker just slashed prices for GPT-5.6 Luna by 80% and GPT-5.6 Terra prices by 20%. Plus, they added a faster option for GPT-5.6 Sol in the API. They're also upgrading Auto-review in the ChatGPT app and Codex CLI from GPT-5.4 to GPT-5.6 Luna, putting a faster, cheaper model behind everyday code reviews. You can explore new plans here.
Anthropic's Claude models broke into real systems during cyber tests: The AI lab just revealed that a misconfigured eval gave Opus 4.7, Mythos 5, and an unreleased internal model live internet access, even though their prompts said the environment was fake. The models mistook the real systems they found for part of the test and compromised three companies using basic tactics like weak passwords and SQL injections. In one case, a poisoned package even tricked a security firm's own scanner into installing it. Check out the postmortem here.
GitHub's new feature simplifies large code changes: The code-hosting giant just shipped stacked pull requests in public preview. This new feature breaks large PRs into small, reviewable pieces, allowing engineering teams to keep big updates moving without sacrificing quality. Best of all, it keeps “main” branch protections intact. Former senior engineer breaks down this entire concept for you to implement.
PRESENTED BY MONGODB
See the architectures Bay Area teams are using to move from AI experiments to production-grade agents.
Memory, state, context, retrieval—all of it, working together on one platform. Leave with patterns you can apply immediately.
Learn from the best, in person. Spend a day at Pier 48 for MongoDB BuildFest learning alongside the architects, engineers, and founders shipping agents into production.
Gain hands-on experience and walk away with patterns you can implement in your own stack.
INSIGHT
Stop chasing the next new model. Here's how to master the harness you already have.
Source: The Code, Superhuman
The next big thing. Every day there's a new MCP, model, skill, or workflow, alongside some viral post claiming one prompt solved it all. Despite the hype, most of it doesn't actually help developers. In a detailed blog, a senior developer at Microsoft argues that the constant chase is counterproductive.
The starting point. By using a workflow that runs exclusively on GitHub Copilot's existing features, he makes the case that mastering the harness beats everything else. The skills and MCPs everyone is racing to install can wait until your workflows actually get complex. Just use your existing tools and build a better harness.
A repeatable workflow is all that matters.
Turn on YOLO mode (Allow All) so the agent runs without pausing for approval at every step. (Disclaimer: Probably best to sandbox it in Codespaces rather than a local machine, since mistakes on private data can be costly).
Start with a prototype to see variations early and surface edge cases before writing real code. A medium-sized model on medium reasoning is more than enough.
Plan methodically, guiding the model with your own expertise instead of accepting every suggestion it makes.
Implement, then review and iterate. The first result will miss, and that's expected.
Refine until the code and UI meet your standards, then run a Rubber Duck review from a different model to catch what one model misses.
Don't settle for good enough. Insisting on quality is still very much “your” job.
Don’t overcomplicate it. Find the simplest repeatable process that gives you high-quality results and learn the harness well. And remember, nothing here is absolute. The “right” workflow today could easily be redundant tomorrow.
P.S: If you’re ready to master harness engineering, this crash course is a good starting point.
IN THE KNOW
What’s trending on socials and headlines
Meme of the day.
Agent Certified: LangChain just dropped an agent engineer certification (at a discount). Complete the courses and pass the proctored exam to earn your credentials.
Tiny Titan: Thinking Machines' new open-weights Inkling-Small handles image, audio, and 1M-token context. You can run it locally (1M views).
Coding GPS: A veteran dev educator just dropped a skill that changes how you plan your most ambitious projects. See how it works (1.9K bookmarks)
Skip Mistakes: An ML engineer with 20+ years of experience just dropped eval skills for Claude Code. It’s built to catch mistakes engineering teams make most often.
Interview Intel: This post shares what OpenAI's software engineering interview process looks like. A secret fifth round tests whether you can code alongside AI (1.9M views).
TOP & TRENDING RESOURCES
Click here to watch the tutorial.
Top Tutorial
7 ways to cut your Claude Code token usage in half: You’ll learn how to cut your token usage in half when using AI coding tools like Claude Code. This tutorial breaks down context engineering and workflow tips to help you simplify prompts, save on API costs, and code way more efficiently with AI.
Top Repo
Claude-Codex Settings: This repo offers a battle-tested coding agent setup with skills, commands, hooks, sub-agents, and MCP servers. It also includes support for the Kimi, MiniMax, and GLM APIs.
Trending Cookbook
How to build agent-first apps: Traditional software typically separates user interfaces from AI, making it difficult to keep human actions and AI agents in sync. This guide shows how unifying UI and agent logic through a shared framework allows specialized agents to run applications and collaborate directly.
AI CODING HACK
How to upgrade your MCP server without hand-editing every import
Anthropic's Model Context Protocol went stateless on July 28, and TypeScript SDK v2 was launched alongside it. Because package names, the tool API, and error types have all been updated, upgrading will break any file that imports the SDK.
The MCP maintainers have outlined the migration path, and the SDK now includes a codemod to handle the renaming automatically. Just be sure to commit your changes first, as the codemod rewrites files in place.
# add v2 alongside v1
npm install @modelcontextprotocol/server @modelcontextprotocol/client
# rewrite imports, .tool() calls, error types
npx @modelcontextprotocol/codemod v1-to-v2 .P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.
IN CASE YOU MISSED IT
Our most-clicked story from yesterday
There is a senior AI engineering role paying $785,000, and it doesn't even require experience in model training. Here is the 10-step roadmap to landing it.
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.
What did you think of today's newsletter?
You can also reply directly to this email if you have suggestions, feedback, or questions.
Until next time — The Code team