Adventures On The Edge

Modern AI systems often blur the words model, agent, and tool. That can make papers about agentic systems difficult to follow. This page is a clarification note for developers learning how these pieces fit together.

Why This Distinction Matters

When paper says, "the agent searches the corpus," it does not usually mean the language model literally opens files, runs commands, or executes code by itself.

More precisely, the language model reasons about what should happen next. The surrounding agent runtime receives that decision, executes the requested tool call, and then returns the result back to the model for further reasoning.

User
  ↓
Agent Runtime
  ↓
LLM / Model
  ↓
Tool Request
  ↓
Tool Execution
  ↓
Corpus / Environment

The Short Version

• The model reasons.
• The agent orchestrates the workflow.
• The tools perform concrete actions.
• The corpus provides the evidence.

What Is the Model?

The model is the language model itself. Examples include GPT, Claude, Gemini, Llama, or another LLM.

The model is responsible for reasoning, interpreting text, forming hypotheses, deciding what information is needed, and producing responses.

However, the model does not directly interact with the operating system. It does not literally run grep, open a file, or inspect a directory. Instead, it produces a structured request that the surrounding system can execute.

What Is the Agent?

The agent is the larger system wrapped around the model.

An agent usually includes:

• the language model
• system instructions
• tool definitions
• tool orchestration
• runtime policies
• memory or working context
• context management
• execution safeguards

In casual language, people often say "the agent searched the corpus" because the agent is the complete operational unit. But internally, the model and runtime are doing different jobs.

What Are Tools?

Tools are callable capabilities exposed to the agent.

In a developer-oriented system, tools might include:

• grep or rg for text search
• find or globbing for file discovery
• file reads
• shell commands
• test runners
• build commands
• lightweight scripts
• database queries
• API calls

The tool performs the concrete action. The model decides when a tool is needed and how to interpret the result.

A More Precise Workflow

When an agent appears to investigate something, the loop usually looks like this:

1. The user asks a question or assigns a task.
2. The model reasons about what information is needed.
3. The model requests a tool call.
4. The agent runtime executes the tool.
5. The tool returns an observation.
6. The model reasons over that observation.
7. The model decides whether to answer or continue investigating.

Model reasons
    ↓
Requests search action
    ↓
Runtime executes tool
    ↓
Tool returns evidence
    ↓
Model inspects evidence
    ↓
Model refines hypothesis
    ↓
Model requests another action

Traditional RAG vs Agentic Search

In a traditional Retrieval-Augmented Generation system, retrieval often happens before the model begins its main reasoning step.

Retriever
  ↓
Top-K Results
  ↓
Model Reasons
  ↓
Answer

This is efficient, but it means the model reasons only over the evidence selected by the retriever.

In an agentic system, search can become part of the reasoning loop itself.

Model reasons
  ↓
Tool search
  ↓
Evidence returned
  ↓
Model revises plan
  ↓
Another tool call
  ↓
More evidence
  ↓
Answer

This is the major conceptual shift. Retrieval is no longer just a preprocessing step. It becomes part of an investigation.

Why DCI Makes This Distinction Important

Direct Corpus Interaction, or DCI, studies what happens when an agent interacts with the raw corpus directly through general-purpose tools instead of relying only on a conventional retriever.

In DCI, the agent may use tools such as:

• grep
• file reads
• shell commands
• lightweight scripts

But the model is still not literally executing those tools itself. The model decides what action would be useful. The agent runtime performs the action. The result is returned to the model as evidence.

Developer Analogy

A traditional RAG system is like giving a developer a prepared incident summary.

That summary may be useful, but it also limits what the developer can see.

A DCI-style agent is more like giving a developer shell access to the repository, logs, and local files.

The developer can search, inspect, refine, and verify. The important difference is not just intelligence. It is access.

Common Source of Confusion

Modern AI writing often uses the word agent as shorthand for the whole system. That is understandable, but it can hide the architecture.

When reading a paper or system description, it helps to ask:

• Is this referring to the model's reasoning?
• Is this referring to the agent runtime's orchestration?
• Is this referring to a tool's execution?
• Is this referring to the corpus or environment being inspected?

Cleaner Terminology

Practical Rule of Thumb

When an AI system appears to "do work," separate the behavior into three questions:

1. Who decided what to do? Usually the model.
2. Who executed the action? Usually the tool through the runtime.
3. Who interpreted the result? Usually the model again.

Key Insight

The model is the reasoning engine. The agent is the operational system. The tools are the hands. The corpus is the environment being investigated.

DCI matters because it changes what the agent can observe and act upon. Instead of limiting the model to a small set of retrieved chunks, the agent gives the model a richer way to investigate the corpus through tools.

That richer interface allows the model to participate in retrieval as an iterative process: search, inspect, reason, refine, and verify.