AgentSky ships one API for every coding agent

Welcome back. We're writing this email with extra excitement today because there are two new releases our team is excited to use internally. First, a stealth startup just dropped the equivalent of OpenRouter for coding agents that lets you run your codebase with multiple coding harnesses. And a new startup is actually paying people to rent out their old MacBooks for AI compute.

Also: How to build LLMs from scratch, Uncle Bob shares a tool for coordinating multiple coding agents, and how to change coding agents without losing context.

AgentSky ships one API for every coding agent: A Silicon Valley-based startup just unveiled the “OpenRouter for coding agents.” Devs can pick a harness and a model, and the platform handles the rest. You can swap agents with a config change without losing history or rebuilding integrations. Dev teams can use its Agent Playground to run tasks across different agents with GitHub-connected tools, then compare cost and results side-by-side to find the best workflow. Check out the API here.

You can put your old idle MacBook to work and earn money: The AI startup Eigen Labs just launched a marketplace that routes AI requests to idle Macs for about half the cost of typical providers. It serves open-weight models like Qwen 3.6, Gemma 4, and GPT-OSS through an OpenAI-compatible endpoint. Mac owners never see your data because the system encrypts all prompts and responses. The company claims Mac owners can earn between $120 and $200 a month. Find more details here. 

ElevenLabs brings voice agents to your terminal: The voice AI lab’s CLI v1 has brought its entire API to the terminal. You can now use CLI commands for text-to-speech, music, dubbing, and voice agents with clean, structured output. It even has built-in skills to teach coding agents how to use it, and it works with the company's MCP for Claude. This means teams can now script updates across entire fleets of voice agents from one place. Try it here.

Claude Code, Cursor, Codex. The output ceiling is insane. But "fix the bug" gets you slop. Detailed context gets you shipping code.

The problem: typing out file references, edge cases, and architecture context 40 times a day is exhausting. So you cut corners. Your agent gets lazy input.

Wispr Flow lets you speak full prompts with syntax-aware dictation. camelCase, snake_case, file names all preserved. Auto-tags files in Cursor and Windsurf.

Used by engineers at OpenAI, Vercel, and Clay. Works on Mac, Windows, iPhone, and Android.

Better prompts in. Better code out. Try Free

An exodus across Silicon Valley. Something is spooking the people who run engineering. Across startups and Big Tech, CTOs, VPEs, and heads of engineering are quitting high-status roles with nothing lined up. Veteran engineer Gergely Orosz says he's never seen this many top-tier leaders walk. In his latest deep dive, he spoke to nearly 20 on career breaks. Six in ten were already on their way out.

A range of reasons. From rapid AI adoption (or the lack of it) to long hours and burnout. Then there’s what Orosz calls “founder mode.” In smaller companies, founders are micromanaging engineering, turning VP and CTO roles into low-ROI gigs. Some leaders left to build their own startups; others were pushed out.

Across the board, these were the most consistent reasons:

It doesn’t stop at the top. When leaders walk, the ripple effect hits the entire team. If you're taking a leadership role, interview the company too: do founders want engineering to change, or just get cheaper? And if your job isn't giving you hands-on AI experience, get it elsewhere.

P.S. Our engineering team put together over 10 practical guides to help you kick-start your journey toward becoming an AI-native engineer. Pass this along to your colleagues today. 

Your non-technical teams are using Claude Cowork all day. Their workflows and agents run locally, making them hard to govern and unusable by the rest of the org.

Major let’s technical and non-technical teams build and share production-ready apps and agents, with guardrails, SSO, RBAC, and audit logs built in.

Book a demo. Clear the queue at major.build.

Harness Engineering with Pi Agent, Deepseek, and Gemini: This tutorial teaches harness engineering using Pi Agent, DeepSeek, and Gemini. You’ll learn how to design structured agentic environments, overcoming LLM performance bottlenecks to create reliable, autonomous coding workflows capable of executing complex, multi-step software development tasks.

Move coding sessions between AI agents: Hit a usage limit halfway through a task? This tool lets you carry the session into another coding agent without starting over. It currently works across Claude Code, Codex, Pi, OpenCode, Copilot CLI, Cursor, and more.

Cut Agent Tokens by 2.7x (using an open harness): Agent costs rising despite fixed model pricing? This walkthrough shows how you can use an open harness to keep context lean using offloading, subagents, code mode, and compaction. Learn exactly where your agent tokens go and how to slash them without hurting performance. 

You kick off a long refactor in Claude Code, then sit watching the terminal. Anthropic's Remote Control fixes that. The session keeps running on your machine, and your phone or browser becomes a window into it.

Your files, MCP servers, and tools stay local. Works on Pro and Max. Team and Enterprise admins must turn it on first.

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

An Anthropic engineer shared the exact skill his team uses to get up to speed on anything fast, and it's easier than you think. 

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