Harness Agents: Local Rust Coding Agent with Persistent Memory

What changed

Harness Agents (ha) is a newly available coding agent developed in Rust, designed to operate locally within a user's terminal. A key feature is its ability to maintain persistent memory, storing conversations, tool results, and learned facts in a local SQLite database. This allows sessions to be resumed, inspected, or continued later, providing a continuous development experience. The agent interacts with users through a terminal-based TUI or a plain line mode. It supports executing single prompts for scripts and CI pipelines with various output formats like text, JSON, or stream-json.

Harness Agents works with code through approval-gated tools. These tools enable reading, searching, globbing, patching, editing, and writing files, as well as running processes and shell commands, and inspecting Git repositories. User approval is required for each action, with options to run once (y), allow the rest of the turn (a), or refuse (n). Permissions can be adjusted via /permissions to select modes like ask, auto-edit, or full-auto, though deny rules still apply.

For research and information gathering, ha includes web browsing capabilities through web_search (integrating with Google via Serper or DuckDuckGo) and web_fetch for retrieving page content and links. It also features a persistent Python REPL powered by IPython, where variables and states can persist across cells and turns. This REPL supports background command execution and parallel execution of child agents. The agent can work towards a defined goal, pausing automatically after a set number of turns or when explicitly instructed, and conversations can be resumed using the /resume command. Memory management is handled by the model itself, storing prompt notes, skills, and subagent specifications, with context ranked for the task at hand. The agent supports extending its functionality through a skill system, allowing integration of custom skills from various directories.

Platform Status

• Windows 10/11 x64: Supported, including CI gates.
• Linux: Pending support; builds and CI jobs run, but it is not yet a supported platform.
• macOS: Not tested.

Quick Start (Windows, PowerShell 7)

Prerequisites include Rust, Git, and PowerShell 7. The installation involves cloning the repository, navigating to the directory, and running scripts/Install-Ha.ps1 for a release build. After installation, ha --version can verify the installation, and ha can be run to open the interactive application. Updates are performed by pulling the latest changes and re-running the installation script.

Why it matters for builders

Harness Agents provides developers with a local, persistent coding assistant that remembers context and past actions, significantly streamlining development workflows. The granular control over AI actions through approval-gated tools allows builders to maintain oversight and ensure code integrity. Furthermore, the extensible skill system and the ability to resume complex tasks offer flexibility and power for custom agent development.

Practical impact

Developers can leverage Harness Agents for tasks ranging from code generation and refactoring to research and debugging, all within their terminal. The persistent memory feature means that complex coding sessions can be paused and resumed without losing context, ideal for long-term projects or intricate problem-solving. Builders can experiment with different AI models and configurations, customize agent behavior with skills, and integrate ha into their CI/CD pipelines for automated tasks. The ability to run ha exec with specific output formats is particularly useful for scripting and automation.

Caveats and source limits

Support for Linux is pending, and macOS has not yet been tested. While the agent supports multiple model providers (DeepSeek, OpenAI, Anthropic, OpenCode), specific model availability and performance may vary. The source does not provide independent benchmark results for the agent's performance or efficiency. Pricing information for API usage is not detailed, as it depends on the chosen model provider. The project is currently listed with 0 stars and 0 forks on GitHub, indicating it is in its early stages of adoption.