Hive is a multi-agent harness that runs colonies of agents. A colony is a group of specialized agents working one business process: a persistent, client-facing lead called the Queen, plus however many worker agents the job needs. You describe the outcome, the Queen does one unit of the work first to prove the path, and a colony then grows around that proven path to run the work at scale.
Underneath is one loop controlling many loops. The Queen is an agent loop, and every worker is a clone of it with the same tools and the same model but its own task, so there is no graph to compile and no orchestration boilerplate. Fan-out is a runtime run_worker call rather than a compiled directed graph, and the colony coordinates through a shared tracker ledger and a persistent task plan instead of a data buffer. Workers write results to the tracker, and the Queen validates them with SQL.
Because there is only one kind of agent, reliability lives in that primitive and every agent inherits it: crash-safe park and resume, budget enforcement, real-time metrics, and out-of-band human oversight. Escalation runs through Sentinel, an account-bound Slack or Telegram channel. The loop parks and persists its state to disk, notifies a person, and resumes exactly where it stopped when they reply, with configurable timeouts and escalation policies.
The framework is Python and needs Python 3.11 or later, with ripgrep required for full search. It uses a uv workspace layout and installs through a quickstart script rather than pip, setting up the core runtime, an MCP tools package, an encrypted credential store, and a default model. Models are reached through LiteLLM-compatible providers, and business systems connect as MCP tools. A JavaScript and TypeScript SDK is on the roadmap.
Features
- Colonies of agents: a Queen spawns worker clones on demand for parallel, long-running work
- One primitive, many loops: orchestration is a runtime fan-out, so there is no workflow graph to wire by hand
- Shared tracker ledger: workers report into a ledger and a persistent task plan that the Queen queries with SQL
- Queen personas: a router picks a persona per outcome, with scoped memory that evolves across runs
- Crash-safe park and resume: agent state persists to disk and the run continues from the point it stopped
- Cost enforcement: spending limits, throttles, and automatic model degradation set at team, agent, or workflow level, with real-time tracking and alerts
- Human in the loop: Sentinel escalates out of band to an account-bound Slack or Telegram channel, so any agent can pause for a person at any point
- Model flexibility: Anthropic, OpenAI, Google Gemini, DeepSeek, Mistral, Groq, OpenRouter, and local models such as Ollama through LiteLLM
- Tool connectivity: MCP integrations for CRM, support, messaging, data, file, and internal API systems, shared by the whole colony
- Compute and browser use: general compute use and browser control through a native extension
- Adaptive agents: reflexion, scoped memory, and skills factored out of proven work into repeatable playbooks
- Windows support: native PowerShell 5.1 or later through quickstart.ps1 and hive.ps1, with WSL optional
