Rizel Scarlett on AI agents, MCP, and the future of dev creativity

Rizel Scarlett, tech lead for open source developer relations at Block, took the virtual stage to talk about something both exciting and cautionary – vibe coding.

In a solo session that felt part technical talk, part cultural commentary, she laid out how AI-powered software creation is evolving – and what developers need to understand to use it wisely.

Rizel started by walking through a simple app she built. 

Its purpose? 

Automate the provisioning of LLM credits to participants at developer events. 

The twist?

 She didn’t write a single line of manual code. 

Instead, she relied on AI to wire it all together using OpenRouter. Just type in an email, and the system spits out a key and step-by-step instructions. Easy, elegant – and created entirely by "vibing."

She wasn’t alone. She highlighted other examples from the community: a flight control simulator and a GTA-style taxi game – all built without traditional coding. 

The concept, originally popularized by Andrej Karpathy, is to “let go of control, trust the AI, and react to what it creates.” Karpathy even jokes that he vibes code while skipping diffs and never typing.

Scarlett, though, offered a more grounded vision.

“We’re not calling it no-code or low-code,” she explained. “We’re calling it vibe coding. And we need to do it responsibly.”

The rise of agentic AI and MCP

A key part of vibe coding’s feasibility is a technology called Model Context Protocol (MCP). 

Rizel broke it down like this: MCP is an open standard that allows AI systems to talk to tools like Figma, databases, and calendars, without you having to copy-paste context between apps.

Think of it like APIs for AI agents. 

It’s the invisible bridge between your prompt and your workflow. That connection is what allows a vibe coder to say, “Build me a campfire scene,” and have an AI do the Blender rendering work without further guidance.

The infrastructure is surprisingly straightforward. MCP operates with a trio: host, client, and server. 

The host is the main AI agent, like Claude Desktop or Goose, that initiates the session. 

The client is the interface layer that facilitates tool calls. 

The server is where the magic happens: a defined set of tools, like “create calendar invite” or “fetch Postgres data,” that the agent can use to perform real tasks.

And yes, Pieces has its own MCP server, Scarlett confirmed.

Real-time demo meets Pixar analogy

To ground the discussion, Rizel ran a live demo using Goose, an open-source MCP client. 

She asked Goose to query a Postgres database for students exceeding their GPA average. 

Goose called on Claude to figure out how to construct the SQL query, retried after a bug, and then returned the results – live.

The entire process, she said, was like a scene from Ratatouille. 

Goose is Alfredo Linguini, the nervous chef who doesn’t really know how to cook. Claude is Remy the rat, hiding under the hat and pulling the strings. 

Together, they serve a perfect plate of functional code.

The risk of “too much vibe”

But not all dishes come out right. Rizel was clear-eyed about the risks of unchecked AI coding.

“There’s a dark side,” she warned. “Creative freedom brings real risk.” 

Developers have introduced bugs, shipped insecure code, and even pushed secrets to production – all while relying on AI agents.

That’s why she’s calling for a redefinition. Vibe coding isn’t reckless automation – it’s “flowing with AI while protecting your project, your team, and your future self.”

She outlined best practices to back that up. F

irst, version control is non-negotiable. “Sometimes Ctrl+Z just won’t save you,” she noted. She once asked an agent to resolve a dependency conflict – and it overhauled her entire app’s functionality instead.

Second, keep asking questions. Don’t switch off your brain because an AI is in the loop. 

Push your agents to explain themselves. 

"Why did you make this change?" is a valid question.

Third, use guardrails.

Goose, for instance, supports gooseignore files, which act like .gitignore but for AI systems. They prevent agents from reading or modifying sensitive directories. 

She also recommended using allow lists to block untrusted MCP servers – a vital step given the rise of slop-squatting, where malicious tools mimic legitimate names.

Finally, choose your tools wisely. 

Some of the best LLMs play nicer with certain agents. “Goose works great with Claude Sonnet 3.5,” she said. 

Others might hallucinate or underperform. 

And not every task calls for full autonomy. 

Goose and similar systems support manual approval, smart approval, or brainstorm-only modes. Know when to engage which.

A developer’s duty

The future of software creation might be conversational and prompt-driven, but Rizel doesn’t believe that gives experienced developers a pass. “Vibe coding isn’t wrong,” she emphasized. “But experienced devs have a responsibility – to set the standard, to define what safe, smart coding looks like.”

In other words, the future might be ambient and agentic. But it still needs oversight, ethics, and yes, documentation.

For developers looking to explore what vibe coding could mean for them, she pointed to tools like Goose and the Pieces MCP server. The door is open. 

The question is: can we walk through it without breaking everything?

In Rizel’s view, we can. But only if we take the long view. Only if we vibe responsibly.