Agent Store is now live — install your first enterprise agent in one click. Explore →
Platform Cloud Hybrid Enterprise Six Engines Orchestration Observability Harness Metaprise LLM
Module 04 · Multi-Agent Orchestration

Nine architectures.
Zero custom orchestration code.

Agents collaborate asynchronously, escalate intelligently, and recover from failures — no need to build custom orchestration logic. Pick the architecture that fits your workflow. Switch between them at runtime.

Select an architecture

01
ReAct Loop
02
Plan-and-Execute Sequential
03
Tree-of-Thought Branching
04
Reflexion Iterative
05
Graph / DAG Parallel
06
State Machine Guarded
07
Multi-Agent Systems Team
08
Tool-Router Dispatch
09
Memory-Driven Adaptive
Switchable at runtime — no redeploy
9
Built-In Architectures
0
Custom Code Required
Runtime
Switchable — No Redeploy
3
Advanced Debugging Tools

Nine ways to orchestrate agents.

Each architecture is a proven multi-agent pattern, built into the platform as a first-class runtime option. No libraries to install, no frameworks to learn. Select an architecture in the Dashboard or SDK and your agents adopt it immediately.

Architecture 01

ReAct

Reason + Act in a tight loop

The agent reasons about the current situation, decides which tool to call, observes the result, then reasons again. Ideal for tasks that require real-time interaction with external systems where the next step depends on the last result.

Customer Service Data Lookup API Chains
Architecture 02

Plan-and-Execute

Plan first, then execute step by step

A planner agent creates a full execution plan upfront. An executor agent works through each step sequentially. If a step fails, the planner revises. Best for structured, multi-step workflows where the full scope is known in advance.

Report Generation Compliance Review Onboarding
Architecture 03

Tree-of-Thought

Explore multiple reasoning paths in parallel

The agent generates multiple candidate reasoning paths simultaneously, evaluates each path's promise, and pursues the most viable branches. Excels at problems requiring creative exploration where the optimal approach isn't obvious upfront.

Strategy Analysis Risk Assessment Complex Decisions
Architecture 04

Reflexion

Learn from mistakes within a single mission

The agent executes, evaluates its own output against success criteria, generates self-critique, and retries with improved reasoning. Each iteration incorporates lessons from previous attempts. Ideal for tasks with clear quality metrics.

Code Generation Document Drafting Quality Improvement
Architecture 05

Graph / DAG

Directed acyclic graphs with parallel branches

Define agent workflows as directed acyclic graphs where independent branches execute in parallel and dependent steps wait for upstream completion. Maximum throughput for complex workflows with clear dependency structures.

ETL Pipelines Multi-Source Research Parallel Processing
Architecture 06

State Machine

Explicit states with guarded transitions

Agents follow formally defined state transitions with guard conditions on every transition. Integrates with the Business State Machine engine to ensure agents cannot skip steps or invent invalid workflow paths. Maximum process control.

Loan Origination Claims Processing Approval Workflows
Architecture 07

Multi-Agent Systems

Specialized agents collaborating on shared goals

Multiple specialized agents divide work by expertise — one researches, one analyzes, one writes, one reviews. A supervisor agent coordinates handoffs. Each agent has its own tools, models, and authority level via AURA's OrgKernel.

Due Diligence Audit Teams Content Pipelines
Architecture 08

Tool-Router

Intelligent routing to the right tool for each step

A lightweight router agent classifies each incoming request and routes it to the appropriate tool or sub-agent. No complex reasoning chains — just fast, accurate dispatch. Best for high-volume scenarios where speed matters more than deep reasoning.

Helpdesk Triage API Gateway Intent Classification
Architecture 09

Memory-Driven

Long-term memory shapes every decision

The agent accumulates context across interactions — conversation history, prior decisions, learned preferences, and organizational patterns. Each new action is informed by the full memory of past interactions. Ideal for long-running relationships and adaptive workflows.

Account Management Personalization Adaptive Workflows
Advanced Feature

Time-Travel Debugging

Rewind any agent execution to any prior state and fork to explore alternative paths. Understand not just what happened, but what would have happened if the agent had made a different decision at any step.

Capabilities
State Replay: Rewind to any checkpoint in a completed or in-progress mission — see exactly what the agent knew and what tools were available at that moment
Fork & Explore: From any historical state, fork a new execution branch to explore what-if scenarios without affecting the original mission
Decision Comparison: Run the same state through different architectures or model versions side-by-side to compare outcomes
Root Cause Analysis: Trace a failure back through the execution history to the exact decision point where things diverged from the expected path
How It Works
Built on the Execution Engine's Checkpoint Persistence — every step is already saved, Time-Travel adds navigation and forking on top
Accessible via the Dashboard UI (visual timeline) or the SDK (programmatic replay)
All forked explorations are logged to the Audit Engine for full traceability
Advanced Feature

Cross-Agent RemoteGraph

A unified interface for calling local and remote agents interchangeably. Build agent workflows that span teams, departments, and deployment boundaries — without changing a single line of orchestration code.

Capabilities
Unified Call Interface: Local agents and remote agents are called with the same API — the orchestration layer handles discovery, routing, and connection management transparently
Cross-Team Collaboration: An agent in the finance team's namespace can invoke a compliance agent in the legal team's namespace — with full OrgKernel authority enforcement
Location Transparent: Agents don't need to know whether their collaborators are in the same process, same cluster, same region, or different cloud — RemoteGraph abstracts the topology
Failover & Retry: Built-in retry policies and failover for remote agent calls — network issues are handled at the infrastructure level, not the application level
Integration
Works with all 9 orchestration architectures — Multi-Agent Systems and Graph/DAG benefit most from RemoteGraph
All cross-agent calls are authorized through AURA's ExecutionToken and logged to AuditChain
Advanced Feature

Nexus Cross-Namespace RPC

Cross-namespace service calls with multi-region, multi-cloud routing. Enterprise-grade infrastructure for organizations running agent workloads across multiple environments, regions, and cloud providers.

Capabilities
Cross-Namespace Calls: Agents in isolated namespaces (different business units, different environments) can invoke each other through controlled, policy-governed channels
Multi-Region Routing: Nexus routes requests to the closest healthy agent instance across regions — minimizing latency while maintaining data sovereignty requirements
Multi-Cloud Support: Agent workloads distributed across AWS, Azure, GCP, and on-premise infrastructure communicate through a single unified RPC layer
Namespace Isolation: Despite cross-namespace communication, each namespace maintains strict data isolation — Nexus enforces AURA's namespace boundaries at the network level
Enterprise Use Cases
Global banks running separate agent namespaces per region for regulatory compliance, with Nexus enabling coordinated cross-border workflows
Healthcare organizations maintaining HIPAA-isolated patient data namespaces while enabling cross-department agent collaboration
Multi-cloud enterprises consolidating agent orchestration across AWS production and Azure disaster recovery

Which architecture fits your workflow?

You don't need to commit upfront. Architectures are switchable at runtime — start with one, observe performance in the Dashboard, and switch if a different pattern fits better. Here's a starting point.

Scenario
"My agent needs to call APIs and react to what comes back."

Use a tight reasoning-action loop where each step depends on the previous result. The agent reasons, acts, observes, and repeats.

→ ReAct
Scenario
"I have a 12-step process that must run in exact order."

Create the plan upfront, then execute each step sequentially. If a step fails, the planner revises without starting over.

→ Plan-and-Execute
Scenario
"I need to evaluate multiple strategies before committing."

Explore multiple reasoning paths in parallel, score each branch, and pursue the most promising approach.

→ Tree-of-Thought
Scenario
"Quality matters more than speed — the output needs to be polished."

Let the agent draft, self-critique, and refine iteratively. Each pass gets better based on explicit quality feedback.

→ Reflexion
Scenario
"Multiple independent tasks can run at the same time."

Model the workflow as a DAG — independent branches execute in parallel, dependent steps wait for upstream completion.

→ Graph / DAG
Scenario
"I need 5+ specialized agents working together on one goal."

Assign each agent a specialty — researcher, analyst, writer, reviewer — with a supervisor coordinating handoffs between them.

→ Multi-Agent Systems

Runtime switchable: Every architecture runs on the same Agent OS runtime. Switch between architectures via the Dashboard or a single SDK parameter — no code changes, no redeployment, no migration. Your agents, integrations, and business rules stay exactly the same.

Orchestrate your first multi-agent workflow today.

Every plan includes all 9 architectures. Start free — no credit card required.