Porting a library to a different language with a sentence (randomlabs.ai)

Slate, an autonomous coding agent, ported the popular browser-use (70.3k stars) Python library to a near 1:1 TypeScript/Node.js replica in a couple hours for about $58.32 using the low-end sonnet model. The run made 583 tool calls across 311 requests, averaging roughly 50k input tokens and 2.5k output tokens per request (totaling ~15.55M input and ~777.5k output tokens). Slate created the entire project structure, converted core modules (utils, DOM, browser, controller, agent subsystems), produced docs, examples and test scaffolding, and iterated in a write-build-verify loop—completing most work autonomously and producing a file tree and tests that mirror the original repo. Initial automated work took ~33 minutes to generate the bulk of files; subsequent automated debugging fixed remaining issues until the port could start from a blank page and perform a Google search. The experiment demonstrates practical capabilities and limits of autonomous agents: they can carry out complex, long-running cross-language ports cheaply and quickly when given high-level direction and a feedback loop, but they still miss edge cases (four files were initially omitted) and encounter architectural mismatches (e.g., MessageManager vs AgentHistoryList) that benefit from guided choices (adapter vs refactor). For the AI/ML community this showcases a cost/time-effective workflow for code migration and large refactors, highlights the importance of human-in-the-loop confirmations, and flags where agent-driven development will need improved verification and architectural reasoning.