Aragora: Control Plane for Multi-Agent Vetted Decisionmaking

# Aragora: Control Plane for Multi-Agent Vetted Decisionmaking

[![Tests](https://github.com/an0mium/aragora/actions/workflows/test.yml/badge.svg)](https://github.com/an0mium/aragora/actions/workflows/test.yml)
[![codecov](https://codecov.io/gh/an0mium/aragora/branch/main/graph/badge.svg)](https://codecov.io/gh/an0mium/aragora)
[![Python 3.10+](https://img.shields.io/badge/python-3.10+-blue.svg)](https://www.python.org/downloads/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

**Aragora is the control plane for multi-agent vetted decisionmaking across organizational knowledge and channels.**

It orchestrates 15+ AI models—Claude, GPT, Gemini, Grok, Mistral, DeepSeek, Qwen, Kimi, and more—to debate your organization's knowledge (documents, databases, APIs) and deliver defensible decisions to any channel (Slack, Teams, Discord, Telegram, WhatsApp, voice).

**Omnivorous by design**: Aragora ingests common document formats (PDF, Office, Markdown, HTML, JSON, CSV), code, images, audio, and video via a broad connector library. Results route automatically to the originating channel—ask questions wherever you are, get answers backed by multi-agent consensus.

**Unlike chatbots**, Aragora builds institutional memory with full audit trails. Vetted decisionmaking is the engine. The product is a defensible decision record.

**Domain**: [aragora.ai](https://aragora.ai) (available)

## Inspiration & Citations

aragora synthesizes ideas from these excellent open-source projects:

### Foundational Inspiration
- **[Stanford Generative Agents](https://github.com/joonspk-research/generative_agents)** - Memory + reflection architecture
- **[ChatArena](https://github.com/chatarena/chatarena)** - Game environments for multi-agent interaction
- **[LLM Multi-Agent Debate](https://github.com/composable-models/llm_multiagent_debate)** - ICML 2024 consensus mechanisms
- **[UniversalBackrooms](https://github.com/scottviteri/UniversalBackrooms)** - Multi-model infinite conversations
- **[Project Sid](https://github.com/altera-al/project-sid)** - Emergent civilization with 1000+ agents

### Borrowed Patterns (MIT/Apache Licensed)

We gratefully acknowledge these projects whose patterns we adapted:

| Project | What We Borrowed | License |
|---------|------------------|---------|
| **[ai-counsel](https://github.com/AI-Counsel/ai-counsel)** | Semantic convergence detection (3-tier fallback: SentenceTransformer → TF-IDF → Jaccard), vote option grouping, per-agent similarity tracking | MIT |
| **[DebateLLM](https://github.com/Tsinghua-MARS-Lab/DebateLLM)** | Agreement intensity modulation (0-10 scale), asymmetric debate roles (affirmative/negative/neutral stances), judge-based termination | Apache 2.0 |
| **[CAMEL-AI](https://github.com/camel-ai/camel)** | Multi-agent orchestration patterns, critic agent design | Apache 2.0 |
| **[CrewAI](https://github.com/joaomdmoura/crewAI)** | Agent role and task patterns | MIT |
| **[AIDO](https://github.com/aido-research/aido)** | Consensus variance tracking (strong/medium/weak classification), reputation-weighted voting concepts | MIT |
| **[claude-flow](https://github.com/ruvnet/claude-flow)** | Adaptive topology switching (diverging→parallel, refining→ring, converged→minimal), YAML agent configuration patterns, hooks system design | MIT |
| **[ccswarm](https://github.com/nwiizo/ccswarm)** | Delegation strategy patterns (content-based, load-balanced, hybrid), channel-based orchestration concepts | MIT |
| **[claude-code-by-agents](https://github.com/baryhuang/claude-code-by-agents)** | Cooperative cancellation tokens with linked parent-child hierarchy, abort control patterns | MIT |
| **[claude-squad](https://github.com/smtg-ai/claude-squad)** | Session lifecycle state machine concepts (pending→running→paused→completed), pause/resume patterns (implemented from scratch due to AGPL-3.0) | AGPL-3.0 (patterns only) |
| **[claude-agent-sdk-demos](https://github.com/anthropics/claude-agent-sdk-demos)** | Official Anthropic subagent patterns, parallel execution idioms | MIT |

See implementations in:
- `aragora/debate/convergence.py` - Semantic convergence detection
- `aragora/debate/orchestrator.py` - Orchestration patterns
- `aragora/debate/topology.py` - Adaptive topology (claude-flow)
- `aragora/debate/session.py` - Session lifecycle (claude-squad patterns)
- `aragora/debate/cancellation.py` - Cancellation tokens (claude-code-by-agents)

## Debate Engine (Dialectic Roots)

The dialectic framing is the **internal engine**; users get adversarial validation outputs (decision receipts, risk heatmaps, dissent trails). If you're here for outcomes, this section is optional background.

Aragora's debate engine draws from Hegelian dialectics:

| Dialectical Concept | Aragora Implementation |
|---------------------|------------------------|
| **Thesis → Antithesis → Synthesis** | Propose → Critique → Revise loop |
| **Aufhebung** (sublation) | Judge s

... (truncated)