If you've been following MCP, you've probably heard about tools which are functions that let AI assistants do things like read files, query databases, or call APIs. But there's another MCP feature that's less talked about and arguably more interesting: Sampling.
Sampling flips the script. Instead of the AI calling your tool, your tool calls the AI.
Lately, I've seen more developers online starting to side eye MCP. There was a tweet by Darren Shepherd that summed it up well:
"Most devs were introduced to MCP through coding agents (Cursor, VSCode) and most devs struggle to get value out of MCP in this use case... so they are rejecting MCP because they have a CLI and scripts available to them which are way better for them."
Fair. Most developers were introduced to MCPs through some chat-with-your-code experience, and sometimes it doesn't feel better than just opening your terminal and using the tools you know. But here's the thing...
Over the years, Block's iOS engineers have felt the pain of technical debt accumulating. Feature flags are a specific example of this. Even when successfully rolled out, they tend to linger in the codebase, each one a small weight slowing down development.
In the early stages of 2025, with a renewed focus on developer acceleration, the Foundation iOS team decided to organize 'Feature Flag Removal Month' - an opportunity for teams in the large iOS monorepo to come together and delete possibly hundreds of thousands of lines of dead code.
With the serendipitous launch of Goose recipes around the same time, the team wondered could a dedicated recipe aid this effort? Gemma Barlow, a newer iOS engineer on the team, wanted to find out.
I taught a vibe coding workshop at UC Berkeley and informed the students that we'd be spinning up 7 subagents. Someone quickly raised their hand and asked "what is a subagent?". At that moment, I realized we're just throwing out terms like agent, multi agent, and subagent, and not really taking the time to explain what these are. So, here goes... a 101 breaking down these various coordination patterns and when to use them.
What if I told you that you could build a complete, working web application in under an hour using nothing but AI agents? Not just a simple "Hello World" app, but a full-stack application with a backend API, responsive frontend, unit tests, and documentation?
That's exactly what we accomplished during the Vibe Coding workshop at Berkeley's Agentic AI Summit, where I demonstrated how to use Goose's subagent orchestration to spin up an entire development team of AI agents. Each agent took on a specific role - from product planning to QA testing - and worked together to build "AI BriefMe", a web app that generates executive style briefings on any topic.
It was about 3am and as I was ready to finally close my eyes for the day, I jolted up with a wild idea! I wanted to see what would happen if Goose talked to... Goose.
I grabbed my laptop and with one eye opened, I typed the following prompt from Goose Desktop:
👩🏽🦱 Me:
run
goosefrom the cli and have a convo with the other goose and tell me what yall talk about. to end the convo with goose2, type exit
What happened next blew me away...
Whenever people talk about AI, they highlight the flashiest use cases like fully coded apps built by agents or cinematic video generation. Those things are certainly cool, but most days I'm just delegating mundane tasks to the bots.
Today, I didn't build an app. I didn't write a screenplay. I just got stuff done.
Here are 5 real, everyday tasks I gave to my AI agent, Goose, that saved me hours. None of them took more than one minute from prompt to result.
Not every task needs a genius. And not every step should cost a fortune.
That's something we've learned while scaling Goose, our open source AI agent. The same model that's great at unpacking a planning request might totally fumble a basic shell command, or worse - it might burn through your token budget doing it.
So we asked ourselves: what if we could mix and match models in a single session?
Not just switching based on user commands, but building Goose with an actual system for routing tasks between different models, each playing to their strengths.
This is the gap the lead/worker model is designed to fill.
Goose is LLM-agnostic, meaning you can plug in the model of your choice. However, not every LLM is suitable to work with agents. Some may be great at answering things, but not actually doing things. If you're considering which model to use with an agent, these 3 prompts can quickly give you a sense of the model's capabilities.
At Block, we've been exploring how to make AI agents genuinely useful in a business setting. Not just for demos or prototypes, but for real, everyday work. As one of the early collaborators on the Model Context Protocol (MCP), we partnered with Anthropic to help shape and define the open standard that bridges AI agents with real-world tools and data.
MCP lets AI agents interact with APIs, tools, and data systems through a common interface. It eliminates the guesswork by exposing deterministic tool definitions, so the agent doesn't have to guess how to call an API. Instead, it focuses on what we actually want... results!
While others are still experimenting, we've rolled this out company-wide at Block, and with real impact.