Edition One

Hi friend! Welcome to the first edition of The Hybrid Engineer.

This newsletter is about what it looks like to work, build, and think with AI.

I’ll cover the tools, workflows, experiments, and lessons I come across while using AI to write code, automate work, build products, and explore where software development and knowledge work are starting to merge.

The “hybrid” part has two meanings: people are crossing old boundaries between technical and non-technical work, and all of us are learning to work in combination with agentic AI systems.

My goal is practical. Every edition should give you something to learn, try, or apply – whether you want to become more valuable in your job, build internal tools, automate boring work, ship side projects, or create your own business.

Thanks for reading,
Robert

Notes: HTML vs. Markdown for agent-human communication

​Thariq from the Claude Code team kicked off a useful debate: should agents keep producing Markdown by default, or should more outputs become HTML artifacts?
​
​Markdown became the default handoff format for human-agent communication: plans, specs, code reviews, implementation outlines. That made sense while human-editability was the main constraint. Markdown is portable, readable raw, cheap to generate, and easy to version.
​
But that constraint is starting to shift: if the agent generates the artifact and the human mostly needs to inspect, navigate, understand, or share it, Markdown might not be the right optimization target. HTML gives artifacts layout, hierarchy, styling, diagrams, tables, interactions, and much more.
​
Markdown is great for text-first artifacts. HTML is better when structure, layout, and visual hierarchy matter.
​
Personally I’ve already felt this shift since Cursor recently shipped its native Canvas feature: React-based artifacts. I use it for one-off dashboards and whenever I want to dive into a specific part of my codebase interactively. They make human comprehension much easier.

There's a trade-off, of course: HTML is more verbose, using more tokens, which is expensive. Markdown might still be the better choice for files that should be human-edited or purely for internal notes to be read by the agent.

​Thariq provides a bunch of interactive examples here. I encourage you to try this for yourself.
The next time you ask an agent for a plan, review, or synthesis, try: "Make this as a self-contained HTML artifact."

Headlines

• Cursor released a long list of new features, including:
  • Multitask Mode is a new way to parallelize work among async subagents in one chat. In my experience, it is a very good way to work on related but separate tasks in one chat pane without much context switching.
  • The Cursor SDK offers a way to programmatically access Cursor's agents and harness in your own software. It's an exciting "platform move" from the company, enabling many use cases. I was lucky to get early access and built a little Chrome extension for your non-technical teammates.​
  • Orchestrate is a plugin/skill that recursively spawns agents that plan, build, and verify work, driving the ability to delegate complex, long-running tasks to Cursor. One big step toward their stated goal of automating coding.

Cursor @cursor_ai Introducing /orchestrate, a skill that recursively spawns agents to tackle your most ambitious tasks with the Cursor SDK. We’ve used it to: - Autoresearch our internal skills, cutting token use by 20% while improving evals - Cut cold start times on our internal backend by 80% 4:58 PM • May 7, 2026 185 Retweets 2263 Likes Read 86 replies

​

• OpenAI's Codex can now do work inside your Chrome browser. It can work in parallel and across multiple tabs without blocking your browser. Since it has access to your sessions, it closes gaps in giving AI access to work that is not yet accessible programmatically via an API or MCP.

OpenAI @OpenAI Codex now works directly in Chrome on macOS and Windows. It’s even better at working with apps and sites in Chrome, and now works in parallel across tabs in the background without taking over your browser. To get started, install the Chrome plugin in the Codex app. 8:8 PM • May 7, 2026 970 Retweets 10584 Likes Read 452 replies

​

Miscellaneous

• ​10 principles for agent-native CLI – If you're working on a CLI and want it to be used by AI agents, this is a good read.
• ​Agents need control flow, not more prompts – It seemed like with Agent Skills the vibes had moved on from an n8n/Zapier-style deterministic definition of workflows, but maybe we're not there yet?
• ​Claude for Microsoft 365 (Word, Excel, PowerPoint) is now generally available on all paid plans. Some are saying it beats Microsoft's own (OpenAI based) Copilot integrations.
• ​Tutorial: Running a local AI model (Qwen 3.5-9B) on an M4 – note that the author admits "it’s nothing like the output of a SOTA model." But that's not the point: not all tasks require SOTA, and being able to get many things done without variable costs, and even offline is a great option to have!