中核オーケストレーション：理解・ルーティング・編集・ゲート

This chapter looks at the core flow from receiving a photo to publishing it on the menu. The flow presented by the speakers does not give everything to one agent. It separates the role that understands the image and chooses its route, the role that edits the image, the role that checks the result, and the final check that can stop publication.

Overview: Move forward safely by separating decisions

The basic data path is as follows. The system understands the photo, uses that result to choose a route, edits it only when needed, and sends it through quality checks before publication. If an intermediate check fails, the system does not continue without limit until it succeeds; it moves to a path that handles the failure.

• Image understanding and routing: Describe the photo and decide whether to enhance it or leave it unchanged.
• Image editing: Edit a photo that has been selected for improvement.
• QA feedback: Check the edited result and tell the editing side about problems. QA means quality assurance.
• Final post-processing QA: Check the result once more before publication after it has passed editing and intermediate QA.

1. Understand the image and choose a route

First, an image-understanding and routing agent looks at the photo. This agent creates a description from multimodal input. Here, multimodal means handling different kinds of information together, such as an image. The presentation says that a large language model (LLM) creates the description and that the description is converted into structured output.

The process can be written as three stages.

1. Provide the photo.
2. The LLM describes the image content and turns the description into structured data in a defined format.
3. Use that data to choose whether to send the photo for enhancement or skip enhancement.

This separation creates an important boundary. Describing or evaluating the image is not the same as deciding which process to call next. The router does not act on the description as plain text; it uses structured information and decision criteria to choose an operational branch. The presentation does not specify the fields in this structured data or the decision values.

2. Skipping is a choice to protect the original, not a failure

When the system decides that a photo does not need improvement, it does not send it to the editor and keeps the original. This is not a failure to improve it. It is a choice that prioritizes safety and faithfulness to the original: it avoids unnecessary compute and also avoids the risk of making an already good image worse through editing.

For example, if a sufficiently clear food photo is mistakenly sent down the enhancement path, compute may increase without improving quality. Editing could also damage the original quality or the way the food appears. For this reason, the router's decision itself must be evaluated.

Figure: The “Agent Orchestration” diagram shows image input passing through image understanding and routing, then moving to either an enhancement path or a skip path. The enhanced result goes through generation and QA feedback. The visible endpoints are “LLM QA” and “Final Scoring / QA”; the diagram does not explicitly show an editing-agent label, a no-publication failure branch, or publication to the menu. See the video around 5:55 for details.

3. The self-correction loop between editing and QA

For a photo that enters the enhancement path, an editing agent creates a result. QA checks that result and sends feedback back to the editing agent when there is a problem. The editing agent can use the initial inputs and this feedback to edit the image again. In other words, editing is not a fixed one-time operation; it is a loop in which the agent uses the check result to correct itself.

This loop has two outcomes.

• If QA passes, the result moves to the next pre-publication check.
• If it continues to fail after repeated attempts, the system does not publish the photo.

The second outcome follows the fail-closed idea. When the enhancement result cannot be confirmed as safe, the system does not force a new image; it chooses the non-publication side. The presentation indicates that editing can be repeated up to a limit, but it does not give the number. Even if this slightly reduces the range of photos that can be enhanced, the system prioritizes not publishing a problematic result.

4. Final post-processing QA decides publication

After editing and intermediate QA, there is also final post-processing and QA. This check is the last gate that decides whether the photo may be shown on the menu. Therefore, passing intermediate QA and being ready for final publication are not the same thing.

The important point is not to combine stage-specific QA into one large decision. Routing checks whether the system chose the appropriate destination. QA during editing provides feedback for correcting the generated result. Final QA checks the post-processed result before publication. Each stage has a role in finding different failures.

What to remember from this chapter

The example in the presentation is a representative, simplified orchestration, not a complete view of the production system. Therefore, the specific schema, prompts, thresholds, and iteration limit cannot be inferred from it. The reliable design claim is that image understanding, routing, editing, QA, and a final publication gate are connected, and that publication is stopped safely when a failure occurs.

Summary

1. Image understanding describes the photo, and the router chooses enhancement or skipping.
2. When the system skips enhancement, it keeps the original and avoids the cost and quality loss of unnecessary editing.
3. The edited result receives QA feedback and self-corrects within a limit.
4. A result that repeatedly fails is not published, and a separate final post-processing QA check confirms publication to the menu.