How AutoDev Remote Agent Enables AI-Powered Coding via GitHub Actions

Overview

AutoDev Remote Agent is an open‑source MCP service that runs on a user’s server (or in a GitHub Action) and provides AI‑assisted GitHub issue analysis, task planning, code generation, architecture design and test‑case creation. It can be integrated into other projects.

Motivation for a Remote Agent

Running the AI backend on a remote server avoids the high compute and cloud costs of a full‑featured IDE. VS Code is used as a lightweight front‑end; the heavy language analysis and refactoring are performed by the remote agent.

Design Principles

The agent follows current AI‑coding trends (2024): RPC services written in Go/Rust, WebView‑based front‑ends for cross‑platform editors, and vector‑based code indexing. Future work anticipates replacing vector search with grep/ripgrep and leveraging stronger LLMs (e.g., Claude) for full‑stack coding on the server.

Implementation

The prototype reuses tooling from AutoDev Sketch and adds GitHub‑specific wrappers (fetch issue, write results). It is distributed via the GitHub Marketplace at https://github.com/marketplace/actions/autodev-remote-agent and the source code resides at https://github.com/unit-mesh/autodev-workbench.

Tool Configuration

[AIAgent] AI Agent initialized with LLM provider: DeepSeek
[AIAgent] Total enhanced tools loaded: 18
Configuration: { maxToolRounds: 3, enableToolChaining: true, toolTimeout: 1200000 }
✅ Agent initialized: DeepSeek (deepseek-chat)
🔧 Tools: directory, read-file, write-file, delete-file, str-replace-editor, analyze-basic-context, search-keywords, grep-search, run-terminal-command, github-get-issue-with-analysis, github-create-new-issue, github-list-repository-issues, project-memory

Round Mechanism

Augment introduces a “Round” mechanism to limit dialogue turns and avoid infinite loops. Each round consists of a tool invocation followed by a summary.

[ToolExecutor] 🎯 Stage 1/1: 2 tools
[ToolExecutor] Round 2: Executing tool: grep-search
[ToolExecutor] Round 2: Executing tool: analyze-basic-context
⚡ Fast execution: grep-search took 33ms
[ToolExecutor] Round 2: Tool grep-search completed in 33ms

Future Plans

Self‑bootstrapping of the Remote Agent.

Expand the toolset with process‑related capabilities.

Introduce a sandbox mechanism for security, including isolated GitHub Action environments and dynamic Action generation.

Sandbox Strategy

When running in a GitHub Action the agent creates a complete code execution environment and can dynamically generate new actions, ensuring isolation from the host runner.

Repository

Source code: https://github.com/unit-mesh/autodev-workbench