// the find
aden-hive/hive
Multi-Agent Harness for Production AI
Hive is a Python multi-agent execution harness aimed at production workloads — state persistence, crash recovery, DAG-based orchestration, and human-in-the-loop controls. It wraps Claude, GPT, and Gemini via LiteLLM and generates agent graphs from natural language goals. Target audience is teams who've already proven out a single-agent prototype and need the plumbing to run it reliably at scale.
1. The harness abstractions are real production concerns: checkpoint-based crash recovery, session isolation, budget enforcement with automatic model degradation, and observable execution via an event bus — not just another thin wrapper around the chat API. 2. The skill system (SKILL.md files, a registry, tool gating) gives a clean extension point for adding capabilities without touching core orchestration. 3. Graph evolution on failure — automatically rebuilding the DAG when an agent node fails rather than just retrying — is a genuinely different approach from most frameworks that just retry the same prompt. 4. Windows is a first-class target with native PowerShell scripts and ripgrep fallback, which most Python agent frameworks ignore.
1. The 'zero setup / no configuration required' claim is contradicted by the uv workspace layout, MCP server config files, credential stores, and a quickstart script that sets up multiple virtual environments — there's real setup here, it's just automated. 2. Documentation lives almost entirely on an external site (docs.adenhq.com) and the repo's own docs folder is mostly guides that reference the external site; if that domain goes dark, the repo is opaque. 3. The 'self-improving agents' and 'graph evolution' capabilities are the core differentiator but there's no architecture doc explaining the mechanism — the README has a Mermaid diagram and the FAQ has a sentence; the actual implementation is buried in the orchestrator package with no explanation of how graph mutations work or what their safety bounds are. 4. At 10k stars with 5.6k forks the contributor pattern looks more viral than sustained — the GitHub Actions show a weekly leaderboard and bounty system that's effective for attention but tends to produce shallow PRs; the CI workflow and PR enforcement rules exist precisely because this became a problem.