Skip to content

Agent harness

Jump to bottom
Thang Chung edited this page Feb 9, 2026 · 3 revisions

Main components

An agent harness is the software scaffolding or "operating system" that wraps around a Large Language Model (LLM) to make it a functional, autonomous agent capable of executing tasks, managing context, and interacting with external tools. While the LLM acts as the brain (reasoning), the harness acts as the body (doing/safety).

The main components of an agent harness include:

  1. Tool Interaction & Execution Engine: this layer connects the model to external applications, APIs, and environments, allowing it to take action.
  • Action Execution: Handles tool calling (e.g., executing code, API calls, web browsing).
  • Tool Sandbox: Provides a secure environment for running tools, such as terminal access or file system management.
  • Error Handling: Manages and recovers from failed tool calls or invalid output from the model.
  1. Context & Memory Management: the harness curates the information presented to the model, acting as its working memory.
  • State Persistence: Stores intermediate steps, conversation history, and variable values to maintain continuity over long-running tasks. Context Engineering: Reduces, compacts, or compresses the prompt to fit within the model's limitations and prevent "context rot".
  1. Orchestration & Planning (Agent Loop): this component manages the logic of how the agent breaks down complex tasks and sequences actions.
  • Goal Decomposition: Breaks high-level user goals into smaller, executable sub-tasks.
  • Sub-agent Coordination: Manages communication and workflow between specialized sub-agents.
  • Lifecycle Management: Handles task initiation, progress monitoring, and finalization (closing resources, saving state).
  1. Human-in-the-Loop (HITL) Controls
  • For safety and reliability, the harness includes mechanisms for human intervention before executing critical or high-risk actions (e.g., deleting files, sending emails, deploying code).
  1. Prompt & Instruction Management: the harness defines the system-level rules, constraints, and instructions that govern the agent's behavior, often utilizing pre-defined prompts.

Summary Table

Component Responsibility
Tools/Action File system, API connectors, browser
Context/Memory Session history, state persistence, summarization
Orchestration Task planning, loop control, sub-agents
Safety/Guardrails Human-in-the-loop, permissioning

Ref: https://share.google/aimode/DOpVzeiO858vZOist

Projects

.

Clone this wiki locally