// Section 2 — Model provider request · 2 MIN READ

[✓] VERIFIED MANUAL ENTRY — This concept has been rewritten from primary sources and is legally cleared for production.

Harness

The client-side application code that drives the model, parses tool calls, maintains conversation logs, and executes command operations on the local machine.

A harness (also called client harness or agent harness) is the application wrapper that coordinates the AI model. Because the model is a frozen, stateless parameters file, the harness must do the active work: managing user inputs, packaging them into Context, calling the API, intercepting Tool Calls, and executing the actual terminal functions or database scripts.

You build or interact with a harness whenever you:

  • Use Claude Code or GitHub Copilot in your terminal.
  • Initialize an OpenAI client package in a server script.
  • Configure the permissions and active directories for a workspace assistant.

Harness Responsibilities

In an agent workflow, the harness acts as the "operating system" for the model, executing key background tasks:

  1. Request Serialization: Rebuilds the conversation log array and formats it as a Model Provider Request.
  2. Tool call Execution: Intercepts the model's intent payload (e.g. {"tool": "read_file"}), checks the Permission Mode, runs the file system operation locally, and sends back the Tool Result.
  3. Error Handling: If a command throws an exception, the harness catches the stack trace and feeds it back as context, allowing the model to self-correct.
  4. Token Counting: Evaluates prompt lengths to trigger Compaction before the context window overflows.
User Input ──► [ Stateful Harness ] ──► Compiles Context ──► Stateless Model
                      ▲                                          │
                      └───────── Runs Tool Call ◄────────────────┘

# AVOID

Do not assume the model has native shell execution capabilities. If a shell command runs, it is because your harness parsed the text, validated the permission, and spawned a local process.

  • Avoid: Blaming the model when a tool execution fails due to local path misconfigurations.
  • Write: Verify that your harness has correctly configured the working directory (Cwd) and access credentials.

# USAGE

Developer A: "The agent says it ran the tests, but the test database is completely untouched." Developer B: "Check the harness console. The model generated a tool call to run the tests, but the harness threw an unhandled exception while spawning the subprocess, so the command was never actually executed."

// SEE_ALSO

ModelThe compiled, frozen parameters that calculate next-token probability distributions. A model is completely stateless and pure.
Agent EngineeringAn architectural design pattern where an LLM is embedded within a stateful loop, allowing it to deconstruct goals, invoke tools, inspect environment feedback, and self-correct.
Tool CallThe structured request generated by a model during inference, specifying a tool name and arguments it wants the client harness to run.
StatefulThe operational design where a client application (harness) maintains a persistent record of messages, files, and variables across multiple stateless model queries.

// SOCRATIC_VALIDATION

Interactive Concept Quiz

QUESTION 1 OF 3SCORE: 0/3

[EDIT_THIS_TERM_ON_GITHUB ↗]