This week's video

I've been using Grok Bot a lot lately and it got me curious what it would take to build something like it.

What products like Grok Bot actually ship is a pattern: authenticated connectors to the apps you already use, a cloud computer that keeps working when your laptop closes, and computer use for the sites that still don't have an API.

In this week's video I walk through recreating that stack myself:

• Mastra as the agent runtime, with memory and a high max-step budget
• Composio MCP for toolkits and OAuth, including a live Notion connect
• Mastra computer use on an E2B desktop sandbox

You'll see a live Notion OAuth connect, then the agent open a real desktop sandbox and work a site itself. That second half is where "teammate" stops being marketing language.

This video is a paid partnership with Mastra.

Watch the video →

Resources mentioned:

• Mastra
• Composio
• Composio × Mastra guide
• E2B
• Grok Bot

What I'd steal from the stack

A few decisions from the walkthrough that still matter if you swap the tools:

Connectors before clever prompts. If the agent can't auth into Notion, Slack, or Gmail, you don't have an assistant. You have a chat window with a persona.

A real computer beats a session that dies with your laptop. Personal agents that matter keep state and keep working when you close the lid.

Computer use needs budget. Driving a desktop takes many steps. A low max-step ceiling makes the agent look broken when it is really just out of runway.

If your personal AI can't do those things, it is still a chat wrapper.

The other half of the Grok Bot story

A couple weeks ago I published the companion angle: connect Grok Bot to your own agents over A2A.

MCP hangs remote tools on your agent. A2A hangs remote agents on your agent. The remote side keeps its own model, prompt, context, and runtime. In that video I wire Grok Bot to Beacon, my Mastra analytics agent on my network. Grok Bot never sees my analytics credentials. It asks. The agent I trust answers.

That one is about containment: what you are willing to hand an outside agent. This week's video is the flip. Recreate the teammate pattern yourself instead of only plugging into someone else's product. Paid partnership with Mastra on both.

Also worth watching

I Built a Renewal Risk Agent with Docusign MCP. Classification stays in code. Judgment sits in two Mastra agents. A human approves before Workflow Builder runs anything. The portfolio in the video is synthetic. Partnered with Docusign. Repo →

Curated Links & Tools

DHH's Rails World 2026 opening keynote. From Austin on Sep 23. DHH on where Rails goes when agents write a lot of the code: language choice, ownership of the stack, and what a one-person team looks like when generation is cheap.

A2A now has an official CLI. Useful next to the A2A video above. From the terminal: a2a card get, a2a send, --stream for live events, plus task lifecycle management. Go, Apache 2.0. Agentic AI Foundation announcement →

Quick Hits

• claude-code-workflows 2026.09.22. agent-ready v1.6.0 adds a migrate mode for bringing existing docs into the agent-ready structure. codebase-readiness v1.9.0 tightens documentation and architecture-clarity scoring. Release notes →
• Mandalore: the part worth keeping. Anthony Costanzo lost the Mac mini running his assistant. What he missed was the accumulated context. He pulled the memory layer out into a local, version-controlled "signet" any agent can read, so the memory lives in a repo he controls instead of behind someone else's API.

Where does your assistant stop today: API-only tools, a persistent computer, or full computer use? Hit reply with yours.

If your team wants this pattern wired into a production stack, tool surface, computer use, approvals included, I help with that through Mastra consulting or book an intro.

Damian