▶ Watch ↗AI Engineer World's Fair 202621:38
Conference affiliation: Managing Director · Rexmore · 2026
Corey Gallon builds AI-native businesses at Rexmore and created chrome-agent, an open-source command-line tool that gives coding agents programmable control of Chrome. His work connects the mechanics of operating software with the checks needed before relying on an agent’s output: inspecting browser events, measuring a proposed fix, reconciling extracted records and independently challenging conclusions.
Corey’s earlier career included serving as a partner and finance leader at an institutional broker-dealer specializing in algorithmic derivatives and equities trading. He subsequently worked in finance analytics at PwC. His contribution to a 2019 finance benchmarking report emphasized connecting analytics to business decisions: an earlier prediction of available working capital, for example, could help treasury improve returns.
His later responsibilities included leading innovation for a consulting firm’s commercial technology business and teaching finance, data analytics and graduate machine learning at Loyola University Chicago. He now builds businesses at Rexmore, a holding company that acquires and develops companies using small operating teams supported by shared AI tools and practices.
In his framework for maintainable agent-built software, Corey assigns product vision, interfaces, acceptance criteria and dependency planning to the engineer. Agents implement small, independently verifiable units against persistent specifications. The engineer’s task includes deciding how the work will fit together and how to establish that each piece works.
Corey’s chrome-agent exposes the Chrome DevTools Protocol through a command-line interface. An agent can inspect page structure, accessibility data, screenshots, logs and network activity while controlling navigation and input. The tool reads protocol information from the running browser, allowing it to reach capabilities beyond those captured in its optional typed bindings. Reusable programs can execute deterministic browser operations without asking a model to choose every intermediate action.
The design supports shared browser sessions: a person can use the browser while an agent watches for errors, or one agent can operate a page while another monitors network traffic. Separate event subscriptions let each participant observe the signals relevant to its task.
His browser-automation methodology follows a sense-act-verify loop: inspect the environment, perform one action and confirm the outcome through an independent signal. It escalates from simple page interactions to trusted browser input and vision-assisted actions when needed. In demonstrations using infrastructure and accounts he controlled, deterministic code handled routine execution while the model interpreted images where judgment was necessary.
Corey applies similar attention to experimental setup when debugging desktop software. To investigate Handy’s memory growth, he built an isolated environment where agents could inject recorded audio, trigger transcription and measure memory without requiring him to operate every test. Runtime instrumentation helped narrow the investigation across Rust, JavaScript and WebKit despite his lack of prior Rust expertise. He discarded a proposed event-rate reduction after measurements showed no useful effect, retaining changes that suppressed unnecessary visualizer events and targeted their delivery.
His open-source Agent Review project packages three levels of review, each aimed at a different failure:
Corey’s account of applying these methods includes an agent-written demand letter whose self-check missed an automatic contract renewal. Independent reviewers caught the clause before the letter was sent. For a year of scanned accounting statements, independent extraction and reconciliation exposed 56 differences and produced records that balanced to the penny.
Review also changed the memory-leak investigation: it exposed a flawed reproduction and an exaggerated claim that the leak had been eliminated before the maintainer merged the corrected change. That case qualifies the earlier debugging success. Instrumentation and a plausible patch were useful, but the test itself and the claimed scope of the fix also needed scrutiny. Corey’s method makes those checks part of delegation, while leaving room for reviewers to be wrong and for a clean review to miss an error.
▶ Watch ↗AI Engineer World's Fair 202621:38
▶ Watch ↗AI Engineer NYC 2026 Online Talks12:25
Loading more talks and highlights…