Corey Gallon is an AI engineer and creator of chrome-agent, an open-source tool that gives coding agents programmable control of Chrome. His work tackles a practical question for AI engineering: how to make autonomous software observable, reliable, and capable of completing tasks beyond a convincing demonstration.
Before moving into AI-native businesses, Gallon was a partner and finance leader at an institutional broker-dealer specializing in algorithmic derivatives and equities trading. He subsequently worked in PricewaterhouseCoopers’ finance analytics and management consulting practice, advised on quantitative modeling, valuation, and credit risk, and later became chief innovation officer of a consulting firm’s commercial technology business. He also taught finance, data analytics, and graduate machine learning at Loyola University Chicago. At AI Engineer World’s Fair 2026, he was affiliated with Rexmore as managing director.
Through Learning Machine, Gallon writes about agent-assisted programming, browser automation, debugging, and the economics of AI development. Several convictions define his approach:
- Human-owned architecture, agent-written implementation. His framework for maintainable agent-built software assigns product vision, interfaces, acceptance criteria, and dependency planning to the engineer. Agents implement small, independently verifiable units, guided by durable specifications instead of improvised prompts.
- Programmable browser control. chrome-agent uses the Chrome DevTools Protocol to inspect page structure, accessibility data, screenshots, logs, and network activity while controlling navigation and input. Gallon favors reusable command-line programs because deterministic browser operations can execute without routing every intermediate action through a model. Human operators and multiple agents can also share a browser session.
- Sense-act-verify. Gallon’s browser-automation methodology requires an agent to inspect its environment, perform one action, and confirm the outcome through an independent signal. It escalates only when necessary: from simple page interactions to trusted browser input, then vision-assisted actions and humanlike cursor movement. His demonstrations, conducted using infrastructure and accounts he controlled, addressed enterprise interfaces, isolated browser boundaries, and visual challenges. Deterministic code handled routine execution; the model interpreted images only when judgment was necessary.
- Instrumented debugging. To repair a memory leak in the Handy transcription application without prior Rust expertise, Gallon built an isolated test harness, added runtime instrumentation, evaluated competing fixes, and validated the result through extended use.
His smaller projects include claude-mv, which preserves Claude Code history when project directories are renamed, and bose-soundtouch, a Python interface for network-connected speakers.
