Enterprise teams know the problem: promise tools break in production. Point solutions need manual orchestration. Context runs out. You re-explain your architecture ten times.
MultiAgent Orchestrator (MAO) solves this with a production-grade multi-agent runtime for real software teams. Hierarchical planning. Persistent memory. Coordination at scale. MAO adds maestro, planner pipeline, hybrid agent, and desktop UI on top of the free Orchestrator base.
Hierarchical Task Decomposition
The HTN planner sits at the top of every MAO workflow. Its job: decompose complex goals into a structured task graph. When the HTN engine can't decompose a step (an unfamiliar operator, a novel domain), an LLM Oracle fallback generates the decomposition and the result feeds back into the planner's learned library.
Give the planner a high-level objective ("migrate authentication to OAuth 2.0" or "refactor payments for multi-currency"). It produces a dependency graph: which tasks block others, which run in parallel, validation criteria per step.
This drives execution order, enforces dependencies, and creates an audit trail of MAO's reasoning before any code changes.
Maestro and the Hybrid Agent
MAO is built around a maestro - a top-level conductor that routes user intents through the planner pipeline, dispatches work to specialist agents, and consolidates output. The hybrid agent is the unit of execution: it can call models from multiple providers (Claude, Ollama, OpenAI, Gemini), choosing the right backend per task instead of being locked to one vendor.
This means a planning step can run on Claude while a bulk-extraction step runs on a local Ollama model, all in the same workflow.
Dynamic Specialist Factory
MAO's dynamic specialist factory creates new agent types on demand for the current task, beyond the 16 base agents. Need a regulatory compliance reviewer, database migration specialist, or performance profiler? The factory synthesizes it with appropriate context and constraints, then removes it after use.
Static libraries are ceilings. Factories are floors.
Episodic and Semantic Memory Layers
MAO's two memory systems layer on top of the KG and code graph.
Episodic memory logs past tasks, decisions, errors, outcomes. Similar tasks retrieve relevant episodes and apply what worked before.
Semantic memory indexes knowledge: codebase, architecture, team conventions, third-party APIs, documentation. This prevents re-discovering patterns on every run.
Result: the system improves over time. Second migration faster than first. Third refactor avoids previous mistakes.
Task-Graph Coordination
Multiple agents need a shared coordination substrate. MAO uses a task graph persisted to WAL-mode SQLite: every node tracks its state, dependencies, and outputs, and the WAL journal makes the graph durable across crashes and parallel writes. Finished sub-task? Post output to the graph node. Waiting on a result? Read the upstream node. Blocker? Flag it; the planner resequences.
This keeps complex workflows coherent across dozens of parallel tasks and lets the team pick up exactly where it left off after any restart.
Tier 2 Workflow Agents
Beyond Claude Code subagents (one prompt, one context) and Tier 1 tool agents, MAO introduces Tier 2 workflow agents: multi-step pipelines that mix deterministic code, LLM steps, DB checkpoints, and structured MCP elicitation prompts. They are code-driven state machines: a Tier 2 agent can survive a crash mid-run, recover from its last checkpoint, and ask the human a structured question via the MCP elicitation channel without ever blocking the rest of the team.
Tauri Desktop UI and REST API
MAO includes a Tauri desktop UI for task management, monitoring, and visualization, plus a REST API for integration.
Your team submits tasks via UI. Your CI/CD submits via API. Both show real-time agent status, task progress, and output.
No glue code needed.
Comprehensive Test Suite
MAO ships with a comprehensive test suite covering HTN planner + Oracle fallback, maestro, hybrid agent harness, specialist factory, memory layers, task graph, Tier 2 workflow steps, API, and coordination logic. Every release passes before shipping.
Pricing
MAO is currently pre-launch. Pricing will be announced at launch. The base VibeCoded Orchestrator that MAO builds on is free and open source under AGPL-3.0 and is available right now — so you can adopt the foundation today and add MAO when it ships.
MAO is the enterprise tier for teams building reliable, auditable AI workflows.
Install the free Orchestrator to start — join the MAO beta when ready.