知能を担うモデル、コンテキスト、ハーネス

This chapter does not treat an agent's "intelligence" as belonging to one model alone. The speaker considers three parts together: the model, the context, and the harness. He focuses on the harness's role in assembling the agent's execution.

1. What does it mean to own intelligence?

By "owning intelligence," the speaker does not mean only training a model from scratch. It means being able to choose the model, manage the context given to it, and control the execution flow through a harness. When you can handle all three, you are less locked into one specific model or one off-the-shelf execution method.

Model: the part that performs inference

A model produces the next response from the input it receives. In this chapter, the speaker does not make the model layer the main topic. He refers to an earlier Fireworks presentation about models, but that presentation is not included in these materials. Therefore, this chapter does not add new claims about model training or performance.

Context: the information the model sees

Context is the information the model sees for its current decision. It is not only the current question. The speaker gives memory, semantic knowledge, and previous conversations as examples of context. Which information the model sees, in what order, and at what time is important.

• Fixed context: Instructions or rules that stay mostly the same each time. For example, an agent's basic role and safety rules belong here.
• Dynamic context: Information that changes from run to run. This includes the current question, the needed parts of previous conversations, and results returned by external systems.

Example (created to explain the idea): For an agent that searches internal data, put the search rules in fixed context. Put the current question and the documents found by the search in dynamic context. Do not insert all historical data every time; select the information needed for the current task.

Harness: the coordinator that makes the three parts work

A harness is not the model itself. It is not merely a list of tools either. It is the outer system that runs the agent: it gathers context, calls the model, acts on external systems, and makes the returned observations available for the next decision.

The flow works as follows. First, the harness puts the context needed for the task into the model's context window. The model then returns a response. If the response instructs an operation on an external system, the harness performs it. It passes the external system's result back to the agent. The model can then use the new information to produce the next response.

Example (created to explain the idea): Suppose the agent needs to check an inventory system. The model chooses an operation to "check the stock of product A." The harness calls the inventory system and receives "5 remaining." It adds that result to the context and lets the model make the next decision. Calling an operation does not automatically end the agent's work.

The speaker also describes the ability to switch models as a benefit of ownership. Without being locked into one model, you can choose the best available model at that time. This does not mean switching is always easy or that the differences disappear. Models may differ in their response patterns and tool handling, so the harness may need adjustment.

Figure 1: In the diagram, "Agent" is shown with "Model" and "Context" inside a "Harness."

This diagram directly supports the speaker's decomposition into three parts: model, context, and harness. The diagram does not show arrows, so it does not explicitly show the coordination flow. Read it as a reference for the relationship among the three elements. Relevant point in the source video (1:16.5)

2. The harness supplies the right information at the right time

The harness's central job is to give the model the context it needs at the time it needs it. Fixed context alone may not contain the information needed for the current task. On the other hand, inserting too much unrelated information can make important information harder for the model to use. The harness therefore gathers dynamic information for the task and assembles the context before calling the model.

Figure 2: A slide asks "What does a harness do?" and explains that it gives the model the right context at the right time for the task.

This slide directly supports the speaker's explanation of context assembly. It presents the harness as a mechanism that supplies information to the model at an appropriate time. However, the diagram does not visibly label fixed context and dynamic context separately. Relevant point in the source video (2:19.5)

3. Why is a harness necessary?

The speaker shows a slide titled "Why do you need a harness?" and lists six needs for agents. From the plan, the confirmed items include connecting an agent to data and managing context during long runs. In other words, a harness supports the whole ongoing task, not only a convenient one-time model call.

Figure 3: A slide titled "Why do you need a harness?" lists six needs for agents.

The slide shows harness requirements, including connecting to data and managing context during long runs. It does not diagram the external-system response or the agent loop itself. Therefore, it is important not to treat execution steps that are not visible in the image as claims made by the image. Relevant point in the source video (2:38.5)

4. Common points of confusion

• A harness is not the model. The model generates a response. The harness coordinates what the model sees, when it is called, and how its response is executed.
• A harness is not a tool list. It does more than provide tools; it handles context and observations around tool calls.
• Ownership does not mean training a model in-house. In this chapter, ownership means choosing the model, managing context, and controlling harness behavior.

The speaker does not make the model layer the main topic and refers to an earlier Fireworks presentation. Its contents are not included here. Also, the exact arrangement of the three parts in the first diagram is described only to the extent that it can be verified from the image. Claims about models or product quality are treated as statements made in this talk, not as facts independently verified here.

5. Summary

An agent's intelligence comes from the combination of model, context, and harness. The model produces a response, the context supplies material for the decision, and the harness coordinates both with external systems. Owning the harness lets you switch models, provide needed information at the right time, and return external results to the next decision. The customization and evaluation discussed in later chapters can also be understood as questions about which of these three parts to improve.