Welcome back. The demand for the newly released Kimi K3 was so high that the Chinese lab had to freeze signups within 48 hours. Some observers think Moonshot doesn’t have enough compute available to meet the rising demand for their new model. Big AI labs like OpenAI and Anthropic are probably safe for now — but for how long? 

Also: A developer's guide to cost-saving model routing, 12 agentic AI projects for your next AI engineering role, and a look inside Coinbase's agent-run engineering.

AI labs are capping their best models: The Chinese AI lab Moonshot just paused new Kimi K3 signups, barely 48 hours after launch. The company said demand pushed its GPUs near capacity. Now it's splitting membership into consumer and coding tiers to ration compute. Anthropic hit the same wall. Claude Fable 5 will be included in all Max and Team Premium plans, but at 50% of the limits. VC Deedy Das argues that Moonshot is short on infra needed to meet the demand.

Hugging Face breach exposes a blind spot in frontier guardrails: The AI startup recently reported a major breach orchestrated by an autonomous AI agent. It gained entry through a poisoned dataset and harvested internal credentials. When the defense team tried to respond, commercial frontier models refused to analyze the exploit code. To keep the attacker data in-house, the team self-hosted China's GLM 5.2 instead. Now, Hugging Face recommends that all defenders maintain their own vetted, self-hosted model.

Alibaba claims its new model trails only Claude Fable 5: The Chinese e-commerce giant just previewed Qwen3.8-Max, their first 2.4 trillion-parameter model capable of handling images, video, and documents alongside text. Alibaba says it ranks second only to Anthropic's Fable 5, though they haven't released benchmarks to back that up yet. Developers can test it now through Qoder at 10% of the standard price, and the company has promised to release open weights soon.

Enterprise AI adoption was initially driven by tokenmaxxing: more users, more prompts and more AI-generated code. But as AI becomes embedded in development workflows, organizations are questioning whether higher usage equals better results.

Discover why teams are shifting to valuemaxxing, focusing on fast delivery, high-quality code, reduced rework and business impact.

Read the full article

Adoption sends an invoice. Coding agents are spreading across teams, and their costs are rising fast. Most teams fix this by rationing. They limit seats or switch to cheaper models, hoping the output quality holds. A senior platform engineer suggests a different playbook for cutting token spend. To really dig into this guide, we first need to look at what's actually driving these costs. 

Billed by the token. Every word the model reads costs money, whether it's useful or not. If you paste a thousand-line log just to find one error, you pay for all thousand lines. If you ask a top-tier model for a simple docstring, you'll pay five times more than a smaller model for the exact same result. In short, not every problem demands the intelligence of Fable 5.

Fixes live in config.You can't expect every engineer to have perfect prompting discipline. The real fix happens at the environment level. Instead of manual effort, his team baked these optimizations into default settings to make them seamless to replicate: 

The playbook is public. Every technique ships with a config you can copy. This includes everything from the output hook to the subagent templates. Combine this with Anthropic's prompt caching, which cuts the cost of repeated context by 90%.

Agents are non-deterministic, stateful, and dependent on models, prompts, and tools that change without warning. Traditional engineering playbooks to build and operate software for decades do not translate.

LangSmith's new guide breaks down the Agentic Operating Model, aligning people, process, and technology to build, deploy, and scale agents.

Get the guide.

How to plan, build, and run loops with Claude Code (by an Anthropic engineer): You’ll learn how an MTS at Anthropic optimizes Claude Code for complex, long-running tasks. You’ll also learn how to use commands like /loop, /goal, and workflows to automate multi-step processes, delegate work to sub-agents, and reduce micromanagement, ultimately improving your productivity.

Vibe Island: This tool puts all your AI coding agents right in your Mac’s notch. You can monitor tasks and approve permissions instantly. Then, jump back to your terminal without breaking your flow. 

Kimi Code CLI (3.9K ⭐): An AI coding agent that runs in your terminal. It reads, edits, and executes code. It uses Moonshot AI to automate complex engineering tasks directly from your command line. 

How to design APIs for agents: APIs designed for humans rely on defaults to hide complexity, which often causes AI agents to hallucinate. To work effectively, agent APIs must instead prioritize explicit field definitions, precise error messages, and raw facts over helpful abstractions.

When Codex goes off track mid-task, both options are frustrating. You can let it finish and ask for a redo, but that bloats your context. Or, you can interrupt the turn and lose all progress.

Thibault Sottiaux, head of Codex at OpenAI, shared a built-in fix for the CLI. Steer mode lets you inject a message at the next thinking step. This happens while the turn is still running. The agent adjusts course instantly without stopping. To use it, enable Steer mode inside any Codex session:

Toggle the "Steer conversation" setting to on. This update saves directly to your config file, so it stays active across all your sessions.

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

Check out this guide to building an AI stack that helps your team ship faster and more reliably.

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