Full-Stack LLM Agent Delivery: Acceptance Criteria Beyond Benchmark Scores

    This article explains why public benchmark scores are not sufficient for evaluating full-stack LLM agent delivery. Using a generated prototype for a low-Earth-orbit travel company with seat selection and flight booking, it examines the visual, interaction, and business layers, outlines a reusable prompt structure, and discusses what “runnable” delivery and acceptance should cover.

    1. What Benchmark Scores Answer

    According to public leaderboards, some recently released open-source models rank highly on benchmarks for difficult reasoning, software engineering practice, and terminal coding. Specific rankings depend on the current data from each evaluation platform; when citing them, record the evaluation version and measurement methodology as well.

    Benchmarks measure the upper bound of a model’s capability on standardized tasks. Real-world business care more about another question: when a complete requirement is handed to the model, can it deliver a runnable result in one pass? Here, “runnable” needs to be broken down: whether pages open, whether the main flow works end to end, whether edge conditions are handled, and whether data can be saved.

    In recent years, the agent pattern has centered on this question. The model does more than complete code snippets; it also performs planning, generation, self-checking, and repair to form a complete workflow. Using a set of full-stack generation cases as an entry point, this article breaks down how this capability is composed and offers methods for evaluation and implementation.

    2. Case Breakdown: Generating a Full-Stack Prototype from a Single Instruction

    2.1 Input Structure and a Typical Requirement

    Reusing generated results requires sufficiently standardized input. A reusable prompt usually includes the following parts:

    1. Role and goal: state what product to build and which users it targets.
    2. Pages and functions: list the pages and the core operations on each page.
    3. Tech stack: specify the frontend framework, backend language, and database type.
    4. Data model: define entities, fields, and relationships.
    5. Business rules: for example, inventory limits and how duplicate submissions are handled.
    6. Acceptance criteria: list the test scenarios that must pass.
    7. Output format: describe the directory structure, run commands, and files to keep.

    A typical requirement is to generate a corporate website for a low-Earth-orbit travel company and add spacecraft seat selection and flight booking. The generated result showed clear characteristics at three levels.

    Visual layer: The page used a full-screen dynamic background to simulate the view from a cabin window, paired with a monospaced-font data panel, status-indicator animations, and a vertical timeline for storytelling. This indicates that the model has a good grasp of common frontend design language, such as CSS animations, parallax scrolling, grid layout, and font pairing.

    Interaction layer: Seat selection was implemented as clickable SVG graphics, and users could select and switch seats. Programmatically generating SVG requires understanding the relationship among coordinates, paths, and event binding, so the completeness of this part needs to be checked separately.

    Business layer: The booking flow included date selection, seat selection, and passenger information entry, with reservation data written to storage. Each flight had a cap of 6 seats and was locked once full. This is demo-level business logic and still has a distance to go before production. The following pseudocode shows the core idea:

    1. def reserve_flight(flight_id, seat_no, passenger):
    2. flight = db.get_flight(flight_id)
    3. if flight.is_locked or seat_taken(flight, seat_no):
    4. raise SeatUnavailableError()
    5. db.create_reservation(fl