Welcome back. Something strange recently came to light that adds a new layer to the AI token discussion. Chinese resellers are offering Claude at a massive 90% discount by using pooled Max accounts and payment fraud. But there is a major catch: they are funding these low prices by harvesting your private logs and reasoning traces. That data is reportedly being sold directly to Chinese labs to train their own AI models. 

Also: Warp's founder maps how your job shifts in six months, a skill that flags which workflows to offload to AI, and why Cloudflare's Workers lead calls agent access broken.

Google plugs computer use to its fastest Gemini model: The search giant just made computer use a built-in tool in Gemini 3.5 Flash. Developers can now deploy agents that can navigate browsers and desktops to handle tedious tasks like continuous software testing. It also includes optional guardrails that prompt users to approve sensitive actions or shut down a job if a prompt injection is detected. Try it via the Gemini API.

OpenAI upgrades its most-used model and debuts its first chip: The ChatGPT maker just pushed an update to GPT-5.5 Instant. It grasps your intent better, handles complex constraints, and sharpens shopping and local searches. It’s live for paid users now. But that's not the only news from OpenAI. They just unveiled Jalapeño, their first custom chip built with Broadcom. It's designed to make running inference across their API both faster and more affordable. 

Notion drops third-party AIs into shared workspaces: The all-in-one workspace just opened a beta that lets external agents work right alongside your team. You can now @-mention Anthropic's Claude or the Cursor code editor just like a teammate to handle coding, bug fixes, or data analysis. Just set their context and permissions, then track every run in real time. It can kick off parallel jobs or schedule them to keep things moving even after you've closed your laptop.

Point it at any repo and it proves what's actually exploitable, not a wall of maybes.

Star & Deploy it on your repo →

Death by a dozen tabs. Senior leadership burnout doesn't usually happen because of one hard project. It happens because your work is scattered across a dozen tools. Your brain becomes the only glue holding it all together. You deal with lost threads and buried action items. You face missed deadlines and constant context-switching. This drains the mental energy you need for high-level thinking.

Put the agents to work. GitHub launched its Copilot app on June 17. Since then, Ashley Willis, GitHub's Senior Director of Developer Relations, has built about 40 automations into her day. These include scheduled prompts that read her calendar, email, messages, and repos, then tell her what needs attention. Her Morning Brief preps every meeting before she opens a tab, and her Commitments Tracker flags promises she made and forgot.

Agents don't do kudos. She keeps a clear distinction. Automation handles the busy work. She focuses on the human side. Her Commitments Tracker flags missed promises, but she personally handles recognition and connection. This discipline is vital. Most people focus on the wrong goals. They build shallow systems that eventually fail. In fact, analysts estimate that over 40% of AI agent projects are abandoned. This usually happens because they try to automate the wrong workflows.

Adapt her receipts. You don't need 40 automations at once. Just open the Copilot app, or any coding agent you run, and ask it to scan your workspace. It will show you where things slip through the cracks. Then, automate your biggest pain point first. Keep the writes gated while you build trust, which GitHub does by default. If you want to take it deeper, Ashley Willis wrote a full guide on the 40 automations she runs and why each one earns its place.

Your next enterprise deal will ask for SSO before it closes. Most teams scramble for weeks. With WorkOS, you ship SSO, SCIM, and audit logs in a sprint - the same stack Anthropic, OpenAI, Cursor, and Perplexity run on.

Start building free

Claude Code usually locks your prompt when it enters plan mode. This can be a pain during large migrations. You end up watching a cursor spin while your machine works.

Anthropic released a new command called /ultraplan. It moves the planning phase to the cloud. This keeps your terminal open and ready. Just run the command with your task. Claude will draft the plan on Opus while you keep coding locally. 

Just swap in your own task. The CLI shows a status indicator. Once it's ready, open the plan in your browser. You can select any section and leave a comment for Claude to revise. Then choose how to run it. You can approve it in the cloud for a pull request.

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

33 CLI tools you need to try: You'll master modern, high-speed alternatives to standard terminal commands for file navigation, network troubleshooting, and managing Docker and Git, saving you valuable time.

Latitude: A platform that helps teams optimize production AI agents by capturing real traffic and identifying key behaviors. It makes it easy to spot recurring failures and turn them into trackable issues for faster resolution. 

No-mistakes (2.6K ⭐): This repo acts as a local proxy that uses AI to validate your code before pushing. It automatically tests your branch and opens a clean pull request once verified. 

Agentic model routing for coding tasks: Current tools for assigning coding tasks to AI models often struggle because they lack real-time performance feedback. By turning the router into a learning agent that executes code and remembers past outcomes, you can significantly improve model selection and cost-efficiency.

Even though many people claim to understand AI, most actually struggle with the basics. To help out, an AI director at Microsoft used hand-drawn visuals to simplify 10 key terms.

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