aden-hive/hive

What it solves

Hive solves the reliability and scalability gap between AI agent prototypes and production business processes. It provides a production-grade harness that handles state management, failure recovery, observability, and human oversight, preventing agents from being simple one-off scripts or fragile agent chains.

How it works

Hive uses a "one loop, many loops" architecture. Instead of a pre-defined graph, it employs a "colony" system: a lead agent (the Queen) pilots the work first to prove the path, systematizes the process into a repeatable playbook, and then spawns worker clones of herself to execute the work in parallel. Coordination is managed through a shared tracker ledger and a persistent task plan, allowing the system to be crash-safe and resume from where it left off.

Who it’s for

It is designed for teams moving AI agents into production environments where uptime, cost enforcement, auditability, and the ability to execute complex, long-running business processes are critical.

Highlights

• Colonies of Agents: Spawns worker clones on demand for parallel execution without needing to wire a workflow graph.
• Production Harness: Includes crash-safe park/resume capabilities, cost enforcement, and real-time observability.
• Human-in-the-Loop: Features "Sentinel," an out-of-band escalation system via Slack/Telegram for human judgment.
• Model Agnostic: Supports 100+ LLM providers via LiteLLM, including local models like Ollama.
• Extensible Tooling: Integrates with business systems via MCP, including native browser use and general compute capabilities.