This chapter examines the foundation of the FDE agent as JD Pruitt describes it. The key is to create one coherent representation of how a company functions. The model must then use that information to produce a business workflow that is neither too detailed nor missing the important points.
会社の働き方を一つの基準で表す
Represent the company's work from one coherent reference
話者がいう「単一の正しい基準（single source of truth）」は、必ず一つのデータベースや一つの文書にすべてを入れるという意味ではありません。会社がどの部署で、誰が、どの順番で、どの仕事に依存して動くのかを、矛盾の少ない形で表すという意味です。これを作ると、エージェントは個別のメモや資料をばらばらに読むのではなく、会社の動きの関係をたどれます。
The speaker's “single source of truth” does not necessarily mean putting everything into one database or one document. It means representing, in a consistent form, which departments and people do which work, in what order, and which tasks depend on other tasks. With this representation, an agent can follow the relationships in the company's operation instead of reading individual notes and documents as disconnected pieces.
大切なのはグラフの考え方であり、保存技術ではない
The graph concept matters more than the storage technology
JD says that this representation uses a dependency graph. A dependency graph treats elements such as tasks or people as points and uses arrows to show when one element must wait for another. The important point is to represent these relationships as a graph. He says that whether the implementation uses a graph database or another storage technology such as Postgres is not the central issue.
例：承認の順番を表す
Example: Representing an approval order
説明を単純にすると、次のような関係です。
To simplify the explanation, the relationship looks like this:
A simple list of people does not show the condition that C must wait for. In a graph, arrows from A to C and from B to C represent C's prerequisites. In other words, the graph represents not only “who is present,” but also “which task enables which other task.”
The speaker explains that many business workflows proceed linearly, while also containing dependencies and cycles. He also mentions violations of a DAG (directed acyclic graph). However, the source does not define in detail how the cycles relate to DAGs here or which structure the system actually adopts. Therefore, this should not be read as a confirmed description of a specific data structure or design.
図：スライド「Building The Agent」は、会社の業務を依存関係グラフで表し、そこからモデルの学習と展開へ進む三つの段階を示しています。左側の図では、要素どうしの前提関係が矢印で見えるため、単なる資料の集合ではなく、業務のつながりを扱うという考え方が伝わります。これは、話者が説明したシステム構築の全体像を示す図です。
Caption: The slide “Building The Agent” shows three stages: representing company operations as a dependency graph, then training and deploying a model. On the left, arrows make the prerequisite relationships between elements visible. This conveys the idea of handling connected business operations rather than a mere collection of documents. The image shows the overall system-building view described by the speaker.
Produce useful output, not only correct information
依存関係を表せても、モデルの出力がそのまま役立つとは限りません。JDが説明する最初のモデル課題は、メモや文書などの文脈から、正規化された業務フロー（normalized process flow）や分析を作ることです。ここでの正規化とは、表現の違う情報を整理し、誰が何をするか、どの順番か、何が前提かを追いやすい形にすることです。
Even a representation of dependencies does not guarantee that the model's output will be useful as-is. The first model problem JD describes is creating a normalized process flow or analysis from context such as notes and documents. Here, normalization means organizing differently expressed information into a form where it is easy to follow who does what, in what order, and on what prerequisites.
According to the speaker, simply giving a large amount of information to a general frontier model can produce output that is too long. Even if it contains a lot of information, the output is hard to use in practice if it cannot select what the client actually cares about. The answer is not to discard every detail either. The model must keep the necessary detail while making the important points clear.
JD explains this difference by comparing it with a strong consultant. A consultant separates information the client considers important from information that can be handled briefly. The goal is to give the model similar selectivity. However, the speaker does not explain which data or mechanism would encode that judgment. Thus, this is a description of the goal, not a presentation of a finished evaluation method.
For this reason, the team is post-training an open-source model to balance detail and clarity. The goal is not to produce longer text. It is to reliably create a workflow that is faithful to the company's context and useful to its readers. The explanation does not provide details about the training data, the exact model, or the evaluation metrics.
次の課題は、必要な文脈を見つけること
The next challenge is finding the necessary context
The challenge so far has been to turn supplied context into a high-quality workflow. But as the company-wide graph grows, the system must first retrieve the part relevant to the question. The next chapter covers how to traverse a knowledge graph to search for the necessary context and the training that supports this search.