An AI agent for unsubscribing from spam (blog.aqnichol.com)

A developer built an LLM-driven "unsubscribe agent" that automates removing you from mailing lists by first locating unsubscribe links in emails (using the Gmail API to fetch subjects, snippets and raw HTML) and then driving a web browser to complete the vendor-specific unsubscribe flows. Link discovery uses a cheap prompt that asks the model to pick an index from a numbered list of links; when that fails the system chunks the email HTML, injects data-index attributes into <a> tags, and prompts the model on those chunks until it finds a candidate link. Once a link is followed, the agent iterates: it sends screenshots and optional HTML summaries to the model, which returns JavaScript to run on the page. Helper primitives (clickText(), scrollDown(), print(), success()/failure()) and rules in the prompt constrain behavior and reduce risky actions. This approach is significant because it shows how LLMs can be stitched into "computer-using" agents for messy, real-world UI tasks that resist rule-based automation. In tests the agent performed well on many simulated flows and real inboxes (found links for 38 domains; 31 successes, 6 failures, 1 timeout), but it still produces false positives (e.g., missed onclick-driven checkbox semantics on PECO) and brittle behavior depending on prompt design. Practical next steps include using function-calling APIs, more robust DOM reasoning, broader simulation suites, and guarding against prompt-injection or adversarial unsubscribe pages.