aden-hive/hive

What it solves

Hive solves the problem of executing complex, repetitive business tasks that require parallel processing and long-term persistence. Unlike traditional agent frameworks that rely on static workflow graphs, Hive allows a "queen" agent to dynamically shape a "colony" of worker agents to accomplish a goal, ensuring that work can be audited, resumed after crashes, and scaled across many similar units of work (e.g., leads, accounts, or documents).

How it works

Hive uses a single execution primitive: the agent loop. A persistent "queen" agent receives a high-level goal, performs an initial "pilot" run to determine the best method, and then creates a "playbook" and a "colony" of worker agents. These workers run in parallel, executing the playbook and writing their results to a shared SQLite tracker on disk. The queen then queries this tracker using SQL to verify the results and provide a final answer.

Who it’s for

It is designed for users who have processes with many similar units of work that need to be done consistently and the same way every time, and for those who need their agents to run for hours or on a schedule while remaining under human oversight.

Highlights

• Dynamic Orchestration: No pre-defined workflow graphs; the queen agent determines the colony structure at runtime.
• Shared State Tracker: Uses a shared SQLite table to track progress, allowing for easy auditing and resuming from where the work stopped.
• Hilbert Browser Bridge: A Chrome extension that allows agents to operate within the user's own logged-in browser sessions.
• Extensible Tooling: Supports built-in tools (shell, web scraping, charts), MCP servers, and reusable "Skills" in the Agent Skills format.
• Resilience: Every agent saves its state to disk, allowing it to resume exactly where it stopped after a crash.
• Model Agnostic: Compatible with any provider supported by LiteLLM, including local models via Ollama.