// 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.

Agent

An autonomous software system that embeds an LLM within a stateful execution loop, enabling it to call tools, interact with files, and iteratively accomplish complex goals.

An agent is a software entity that uses a language model as its central reasoning engine. Unlike a simple text generator, an agent has agency: it can evaluate a goal, inspect its Environment, choose which Tools to execute, read the results, and adjust its plan dynamically until the task is complete.

You interact with an agent when you:

  • Use Claude Code to run automated repository refactoring.
  • Setup an auto-GPT bot that browses the web to aggregate research.
  • Deploy background loops to automatically audit incoming pull requests.

Agent Architecture

An agent is not a single model file. It is a system composed of:

  1. The Brain (Model): Exposes reasoning and generates action intents.
  2. The Body (Harness): Executes the actual actions, sanitizes inputs, and feeds back logs.
  3. The Senses (Tools): Connections to filesystems, compilers, and APIs.
  4. Memory: Active Context (short-term) and databases/files (long-term).
+-----------------------------------------------------+
|                  AGENT SYSTEM                       |
|                                                     |
|           [ Brain (LLM) ] ◄──► [ Memory ]           |
|                 │                                   |
|                 ▼ (Decides Tool Call)               |
|      [ Body (Harness) ] ──► [ Senses (Tools) ]      |
+-----------------------------------------------------+

# AVOID

Do not treat an agent as a human programmer who possesses global intuition. If the agent makes a mistake, it cannot correct it unless the tool results feed back the compiler errors.

  • Avoid: Letting an agent run without automated test commands mounted, expecting it to know if its code compiled.
  • Write: Always provide validation tools (linters, test suites) so the agent can self-correct when checking outputs.

# USAGE

Developer A: "Our agent wrote the code but didn't verify if it runs." Developer B: "That's because we didn't give the agent a shell tool. It has no way to run the tests. Let's add a test runner tool to its environment so it can execute checks autonomously."

// SEE_ALSO

ModelThe compiled, frozen parameters that calculate next-token probability distributions. A model is completely stateless and pure.
HarnessThe client-side application code that drives the model, parses tool calls, maintains conversation logs, and executes command operations on the local machine.
SubagentA secondary agent spawned by a parent agent to execute a specific sub-task in a separate, isolated context window, returning a brief summary result.
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.

// SOCRATIC_VALIDATION

Interactive Concept Quiz

QUESTION 1 OF 3SCORE: 0/3

[EDIT_THIS_TERM_ON_GITHUB ↗]