HKUDS/DeepCode

"DeepCode: Open Agentic Coding (Agent Harness & Loop Engineering & Multi-Agent Orchestration)"

What it solves

DeepCode is a general-purpose coding agent framework designed to handle real-world software repository work. It moves beyond simple code generation by enabling AI agents to explore codebases, edit files, execute commands and tests, and iteratively refine their work until a task is successfully verified.

How it works

DeepCode employs a multi-agent system utilizing "Loop Engineering" to maintain progress through a cycle of understanding, implementation, verification, and repair. It features an Agent Harness that manages the runtime, allowing users to interact with agents via an interactive CLI or a Tauri-based Desktop workbench. The system supports a flexible Skill system where reusable expertise (Skills) can be discovered from project-specific or personal collections, and integrates with Model Context Protocol (MCP) servers for extended tool capabilities.

Who it’s for

It is built for developers and researchers who need an agentic workflow for building, fixing, and improving production-ready code, as well as those looking to transform research papers into implementable code (Paper2Code).

Highlights

  • Multi-Interface Support: Offers both a terminal-based CLI and a visual Desktop workbench for managing sessions, goals, and code changes.
  • Loop Engineering: Enables agents to iteratively work toward a goal through continuous verification and repair rather than providing a single response.
  • Extensible Skill System: Supports reusable, versioned Agent Skills and MCP server integration for expanded toolsets.
  • Durable Local Work: Maintains session history and project state locally, allowing users to stop, resume, or revise goals without losing progress.
  • Subagent Runtime: Ability to delegate tasks to external backends like Codex or Claude Code as subagents.

Related

  • Project
    deepagent-ltd/deepagent-code

    DeepAgent Code is a self‑hosted AI coding workspace that combines a large‑language‑model runtime with durable, versioned project memory, live planning, multi‑agent collaboration, and fine‑grained human control. It offers three collaboration modes (Auto, Loop, Design), integrates LSP‑based code intelligence, plug‑in domain packs, and an expert‑panel review system, and works with any of 75+ model providers. Install via a one‑line script, configure providers through a UI or CLI, and run tasks like `deepagent-code run "add rate limiting to /api/users"`. The code is AGPL‑3.0‑or‑later and built on Bun.

  • Project
    AMAP-ML/LongHorizon-Harness

    A loop engineering framework that turns existing AI agents into long-running computer-use systems by wrapping them in a structured plan-act-verify cycle across GUI and CLI environments.

  • Project
    anomalyco/opencode

    An open-source AI coding agent available as a CLI or desktop app that helps developers plan, explore, and build software.

  • Project
    Kritt-ai/open-kritt

    An open-source security research platform that orchestrates AI agents to find, validate, and rank vulnerabilities in codebases using focused, parallelized workflows.

  • Project
    vercel-labs/deepsec

    An agent-powered vulnerability scanner for large-scale repositories that uses AI to find and verify deep-seated security issues.