Autonomous Coding Agent

An autonomous AI agent that accepts natural language coding requests via an interactive CLI and produces complete, production-ready Python projects — including source files, a README, and a requirements manifest. The agent plans, writes, reviews, and fixes code end-to-end before handing off a runnable project to the user.

Architecture

The agent is built on the deepagents framework, which wraps LangGraph's stateful graph execution into a higher-level agentic API. A single main agent orchestrates the full workflow, delegating quality control to a specialized code-reviewer subagent. A FilesystemBackend exposes file I/O tools (write_file, edit_file, read_file, ls) to both agents, with all output scoped to a ./projects/ directory.

• Main Agent — receives user requests, runs the Plan → Write → Review → Fix → Deliver workflow
• code-reviewer subagent — reads the generated files and produces a structured PASS / NEEDS CHANGES report; the main agent applies fixes before delivery
• name_project tool — sanitizes the user's request into a lowercase hyphenated slug and creates the project folder

Key Features

• End-to-end code generation — from a single sentence, the agent names the project, plans implementation steps, writes all source files with docstrings and type hints, creates a README with setup and run instructions, and produces a requirements.txt when third-party packages are needed
• Built-in code review loop — a code-reviewer subagent checks correctness, edge cases, PEP 8 style, type annotations, and README accuracy; the main agent applies every flagged fix before delivering
• Structured workflow via todos — the agent uses write_todos to plan and tracks its own progress through each step, ensuring nothing is skipped
• Thread-isolated memory — each task gets a unique thread_id backed by MemorySaver, enabling multi-turn follow-up within the same session
• Real-time streaming — agent steps are streamed via agent.stream() and pretty-printed as they arrive, giving visibility into planning, writing, and review stages
• Interactive REPL — a persistent CLI loop accepts unlimited tasks per session, with an ASCII banner and clear status output

Tech Stack

• deepagents — high-level agentic framework (subagents, backends, tool delegation)
• LangGraph — stateful graph execution and MemorySaver checkpointing
• LangChain OpenAI / GPT-4o-mini — language model backend
• FilesystemBackend — virtual filesystem layer for safe, scoped file I/O
• Python — implementation language