Agentic AI in the Wild brought builders, founders, and practitioners to Cambridge on August 18 for an evening focused on a question many teams are already dealing with: how do you make AI agents work reliably, day after day, on real codebases?

The meetup covered different parts of the agent workflow, from local validation before costly cloud runs, to machine-readable guardrails for AI-generated code, to large codebase migration with agent loops, and a live look at building a working AR scene through React Native. The focus stayed on execution, verification, and quality control.

Agenda

5:45 PM Doors open, pizza and beers

6:10 PM Intro

6:40 PM Prompt to Portal: How AI Makes Spatial Computing Accessible. / Oliver Edis, ReactVision

7:00 PM How to Embed Code Quality into SDLC / Nnenna Ndukwe, Qodo

7:15 PM - 7:55 PM Panel discussion

8:00 PM - 8:15 PM Fuzz First, Farm Later: Local Validation Before Costly Cloud / Matt Hargett, Callstack

8:15 PM - 8:30 PM Large Codebase Migration with Loop Engineering / Rafael Mendiola, okthink

7:35 PM Networking and pizza

Talks and speakers

Prompt to Portal: How AI Makes Spatial Computing Accessible

This talk showed an AI agent building a working AR scene live on stage through an MCP server, with no Unity and no new engine, just the React Native stack. The session also covered what the agent has access to under the hood and why building for physical space differs from building for a screen.

Oliver Edis - ReactVision

Large Codebase Migration with Loop Engineering

A large React app migration from Vite to Expo brings a lot of moving parts with it: components, styling, browser APIs, and tests all need to shift without breaking the product. This talk looks at a different way to handle that kind of work. Instead of a big project plan split into dozens of tickets, the team built an Agent Loop: a coding agent re-prompted every 30 minutes, finding one manageable piece of the migration, converting it, running tests, recording progress, and leaving the code ready for the next run.

Rafael Mendiola – okthink

How to Embed Code Quality into SDLC

As AI tools generate more code, weak standards get exposed fast. This talk looked at how teams can codify development best practices into machine-readable guardrails so AI-generated code respects a system’s design.

Nnenna Ndukwe walked through context engineering at scale and a practical methodology for embedding code quality into the SDLC using Codex and other AI dev tools, applied where developers already operate: CLI, IDE, and Git.

Nnenna Ndukwe – Qodo

Fuzz First, Farm Later: Local Validation Before Costly Cloud

Agent-driven testing on real device clouds closes an important part of the verification loop, and the bill scales with your commit rate. This talk covered what you can catch before you touch a device: branch mutation testing, coverage-guided fuzzing, and deep CPU-level profiling run on a laptop’s efficiency cores as a stand-in for constrained phone, watch, XR, or TV silicon.

Matt Hargett (Callstack)