Quick Guide to LangChain DeepAgents: Exploring the Production‑Grade DeepAgents Code Framework

Project Overview

DeepAgents Code (dcode) is an open‑source terminal coding agent located at

https://github.com/langchain-ai/deepagents/tree/main/libs/code

. It is built on top of the DeepAgents SDK and serves as a reference implementation of a production‑grade agent.

Architecture

The repository follows a client‑server (C/S) design where the UI (client) and the agent runtime (server) run in independent processes. This separation prevents UI blocking when long‑running shell commands are executed, avoids tangled state management, and enables future extensions such as remote execution or multi‑user sharing. Communication between client and server uses a streaming protocol based on LangGraph’s astream_events API; each event carries a type (model output, tool call, state update) and payload that the client consumes to update the UI.

Code Layout

Key modules are grouped by their runtime side:

Server (agent core) : main.py, agent.py, server_graph.py, server.py, server_manager.py Server (tool system) : tools.py, mcp_tools.py, mcp_auth.py, mcp_commands.py, mcp_oauth_ui.py, mcp_login_service.py, mcp_trust.py, mcp_providers/ Server (state & memory) : sessions.py, resume_state.py, offload.py, memory_guard.py Server (skills & sub‑agents) : skills/, subagents.py, built_in_skills/ Shared (configuration & context) : config.py, configurable_model.py, local_context.py, model_config.py, project_utils.py Server (sandbox integration) : integrations/ Client (Textual application) : app.py, app.tcss, editor.py, input.py, output.py Client (UI components) : widgets/ Client (command system) : command_registry.py, config_commands.py, auth_display.py Client (terminal capabilities) : terminal_capabilities.py, terminal_escape.py, unicode_security.py, itern_cursor_guide.py Shared (infrastructure) : _constants.py, _env_vars.py, _paths.py, _version.py, _cli_context.py, _session_stats.py, _startup_error.py, _debug.py, event_bus.py, doctor.py,

extras_info.py

Technology Stack

DeepAgents SDK – Provides create_deep_agent to assemble models, middleware, and tools, and supplies state management and sandbox support.

Textual + Rich – Textual builds the interactive TUI; its reactive attributes automatically refresh the UI on state changes, which is essential for token‑by‑token streaming. The main app class in app.py inherits from textual.app.App, and app.tcss defines terminal CSS. Rich handles rendering, syntax highlighting, markdown, progress bars, and tables.

SQLite + aiosqlite – Persistent session data is stored in a single‑file SQLite database. The aiosqlite wrapper runs DB operations in a background thread to keep the UI responsive. For web‑scale deployments the author recommends PostgreSQL or MySQL.

Streaming protocol – Client and server exchange events via LangGraph’s astream_events API. Each event includes a type (model output, tool call, state update) and payload, which the client parses to update the UI.

Local file strategy – Configuration uses TOML. Long‑term memory, skill definitions, sub‑agent definitions, and system prompts are stored as Markdown files (e.g., AGENTS.md, SKILL.md, AGENT.md, system_prompt.md), implementing a “file‑as‑database” approach that lowers the barrier for user customization.

File‑Based Data Layout

Long‑term memory: AGENTS.md stored under ~/.deepagents/<agent>/ and .deepagents/ Skill definition: SKILL.md stored under ~/.deepagents/<agent>/skills/ and .deepagents/skills/ Sub‑agent definition: AGENT.md stored under .deepagents/agents/ System prompt: system_prompt.md packaged as a resource within the repository