Workflow Engines & Orchestration – Obot

Track: Sessions – Day 1 | Ref: 1.7 Speaker: Donnie Adams, Software Architect, Obot

Overview

Obot presented workflow engine patterns for MCP tool orchestration, covering how event-driven architectures, state machines, and DAG-based workflows can coordinate agents and compose MCP tools into reliable production systems.

Workflow Engine Patterns

Pattern	Description
Event-driven architectures	Workflows triggered and coordinated by events rather than polling
State machines	Explicit state transitions governing agent behavior and tool invocation
DAG workflows	Directed acyclic graphs defining dependency-ordered task execution
Long-running process management	Durable workflows that survive restarts and handle multi-step operations

MCP Integration

• Tool composition: Combining multiple MCP tools into higher-order operations
• Workflow-as-code: Defining orchestration logic programmatically rather than through configuration
• Agent orchestration layers: Middleware that routes requests to appropriate agents and tools
• Error handling: Retry policies, circuit breakers, and fallback strategies for MCP tool failures

Agent Orchestration

• Multi-agent coordination: Multiple agents collaborating on a shared task
• Task delegation: Routing subtasks to specialized agents based on capability
• Parallel vs sequential execution: Optimizing throughput by parallelizing independent steps
• Result aggregation: Combining outputs from multiple agents into a coherent response

Coordination Patterns

Pattern	Use Case
Centralized	Single orchestrator manages all agents and tool invocations
Distributed	Agents coordinate peer-to-peer without a central controller
Hierarchical	Parent agents delegate to child agents in a tree structure
Event-sourced	All state changes captured as events for replay and audit

Architecture

┌─────────────────────────────────┐
│      Orchestration Engine       │
│  (State machine, DAG, events)   │
└──────┬──────────┬───────┬───────┘
       │          │       │
       ▼          ▼       ▼
┌──────────┐ ┌────────┐ ┌────────┐
│  Agent A │ │Agent B │ │Agent C │
│ (Planning)│ │(Coding)│ │(Review)│
└─────┬────┘ └───┬────┘ └───┬────┘
      │          │           │
      ▼          ▼           ▼
┌──────────┐ ┌────────┐ ┌────────┐
│ MCP Tools│ │MCP Tools│ │MCP Tools│
│ (Search, │ │(Code,  │ │(Lint,  │
│  Plan)   │ │ Build) │ │ Test)  │
└──────────┘ └────────┘ └────────┘

Key Takeaways

• Stateless design: Workflow steps should be stateless with externalized state for durability
• Idempotency: Every MCP tool invocation should be safe to retry
• Observability: Instrument every step for tracing, logging, and metrics
• Modularity: Compose small, focused tools rather than building monolithic agents
• Error handling: Plan for failure at every step — retries, timeouts, fallbacks, and circuit breakers