Plan-then-execute architecture slashes browser-agent costs 50×

A developer running browser automation for AI agents — handling sign-ups, form fills, OTP extraction, and verification flows — cut LLM costs by 50× after abandoning the standard continuous agent loop used by tools like browser-use and Stagehand. The continuous loop architecture is the culprit: every action step re-sends the full accumulated page context to the model, meaning token counts compound across each interaction. The replacement is a two-phase approach where the model generates a deterministic execution plan upfront, then a lightweight executor runs it step-by-step without further LLM calls for each action. For structured, predictable web interactions, this collapses the per-task cost dramatically. Essentially: one anonymous developer, browser-use/Stagehand as the incumbent tooling, and a plan-then-execute pattern as the challenger. - The original loop architecture sent full page context to the LLM on every action, not just once per task. - Two-phase separation means LLM cost is paid once at planning time, not N times across N actions. - The post includes specific token counts and cost breakdowns comparing both architectures directly. For teams scaling browser agents in production, this is less an optimization trick and more a fundamental rethink of where the LLM sits in the execution loop.