Coding Agent Architecture: How AI Coding Tools Work Under the Hood

Coding Agents: full architecture breakdowns

• t3code Architecture — react, astro, SQLite
  Minimal web GUI for coding agents by Theo
• opencode-dev Architecture — React, Node.js
  OpenCode
• codex Architecture — Rust
  Codex CLI Coding Agent
• claude-code Architecture — Rust
  Cluade Code Coding Agent
• cline Architecture — JavaScript, Node.js, TypeScript
  Coding Assistant

Patterns that show up in every coding agent

The agent loop

The model proposes an action, the harness executes it, and the result is fed back into context for the next turn. Every agent we analyzed implements this loop — the differences are in how they decide when to stop, how they handle failures mid-loop, and how much autonomy a single turn is given.

Tool dispatch layer

File reads, edits, shell execution, and search are exposed to the model as first-class tools with schemas. The dispatch layer that routes model output to real system calls is where most of the engineering lives — and where the safety checks sit.

Context management

A repository does not fit in a context window. Repo maps, file summaries, selective reads, and truncation strategies are how each agent decides what the model gets to see. This is the layer that most differentiates a usable agent from a lost one.

Permissioning and sandboxing

From simple allow-lists to fully sandboxed execution environments, every agent draws the line between 'the model suggests' and 'the system executes' differently. This is where the architectures diverge the most across the projects below.

More agent system architectures

• openhuman Architecture — Bash, rust
  Agent that sees, hears and learns
• hermes-agent Architecture — Python, React
  Hermes Agent
• clawdbot Architecture — TypeScript, Node.js, Docker
  AI Assistant
• AgentGPT Architecture — Unknown
  Autonomous agent system in browser
• babyagi Architecture — Python
  Autonomous task-driven AI agent implementation

Frequently asked questions

What is a coding agent?

A coding agent is a system that wraps a large language model with tools for reading files, editing code, and running commands, plus a control loop that lets the model work toward a goal across multiple steps instead of answering a single prompt. Claude Code, Codex, Cline, and OpenCode are all coding agents with openly inspectable architectures.

How does a coding agent execute code safely?

Through a permission and sandboxing layer between the model and the operating system. The model never executes anything directly — it requests a tool call, and the agent harness validates it against permission rules, sandbox boundaries, or user confirmation before running it. Each project on this page implements that boundary differently.

How do coding agents fit large repositories into the model's context window?

They don't — they select. Techniques include repository maps (compressed structural overviews), on-demand file reads, search tools, and summarization of prior conversation turns. The context management module of each agent decides what the model sees at every step.

What is the architectural difference between Claude Code, Codex, and Cline?

The core loop is similar; the differences are in packaging and boundaries. Claude Code and Codex are terminal-first CLI agents, while Cline lives inside VS Code as an extension and leans on the editor for file operations. Their tool sets, permission models, and context strategies each reflect that host environment — the individual analyses below map the exact modules.

Can I get this analysis for my own codebase?

Yes — these pages are generated by Revibe's codebase analysis. Upload a repository or import it from GitHub and you get the same architecture diagrams, module maps, and execution flows for your own project.