DISCOVER THE FUTURE OF AI AGENTSarrow_forward

Quoracle

calendar_todayAdded Apr 23, 2026
categoryAgent & Tooling
codeOpen Source
Workflow Automation大语言模型Multi-Agent SystemModel Context ProtocolAI AgentsAgent FrameworkWeb ApplicationAgent & ToolingModel & Inference FrameworkAutomation, Workflow & RPAEducation & Research ResourcesProtocol, API & Integration

Recursive agent orchestration with multi-LLM consensus, enabling governed large-scale agent collaboration via multi-model voting and tree-structured delegation.

Core Mechanism#

Quoracle is a recursive multi-LLM consensus agent orchestration system built on Elixir/Phoenix. Every agent decision must go through multi-model voting from a model pool before execution, supporting 0–9 refinement rounds (default 4, temperature auto-decreases per round), using embedding models for semantic similarity comparison to aid consensus.

Recursive Hierarchical Agents#

Agents can recursively spawn child agents forming tree-structured hierarchies with Sequential or Parallel delegation strategies, each maintaining independent per-model conversation history. Each task has its own PubSub topic to prevent cross-task state leakage.

Grove Orchestration System#

Machine-readable GROVE.md manifests declare complete agent topologies, governance rules, filesystem isolation, JSON Schema validation, and bootstrap configuration—analogous to Docker Compose for containers. Includes built-in MMLU-Pro and LiveBench benchmark groves ready to use.

Governance & Security#

  • shell_pattern_block: Pattern-based blocking of dangerous shell commands
  • action_block: Blocking by action type
  • confinement: Filesystem path restrictions
  • JSON Schema validation: Write operations must pass Schema checks
  • NO_EXECUTE marking: External content auto-tagged as non-executable for injection prevention
  • API key encryption: AES-256-GCM encrypted storage

Note: README explicitly states "not for unsupervised production deployments"—no user authentication, no shell sandbox.

Skills System#

Reusable SKILL.md files (YAML frontmatter + Markdown) injected into agent system prompts, with learn_skills action for runtime dynamic skill acquisition.

Multi-LLM Providers#

Unified access via ReqLLM to cloud models (OpenAI, Anthropic, Google, Azure OpenAI, Groq) and local/self-hosted models (Ollama, vLLM, LM Studio). MCP tool integration via anubis_mcp (~> 0.17.0).

Real-time Dashboard & Persistence#

Phoenix LiveView browser dashboard provides real-time visibility into all agent states (consensus rounds, messages, hierarchy), fully persisted via Ecto + PostgreSQL (≥ 14). Frontend built with Tailwind CSS + esbuild.

Deployment#

  • Source development: Requires Elixir ≥ 1.18, OTP ≥ 27, PostgreSQL ≥ 14, libvips
  • Docker: docker-compose up
  • Release Tarball: Self-contained package (Unix only), no Elixir/Erlang needed

Current version 0.2.4, AGPL-3.0.

Related Projects

View All arrow_forward

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.

rocket_launch