Multi-tool orchestration under uncertainty for enterprise automation

Abstract

Enterprise automation requires agents to coordinate heterogeneous tools — browsers, APIs, email clients, and code execution environments — under conditions of partial observability, unreliable tool responses, and evolving application interfaces. We describe an orchestration framework built on observe-orient-decide-act (OODA) cycles with error-driven re-planning. The system constructs execution plans as directed acyclic graphs with parallel and sequential dependencies, monitors each node for failure signals, and triggers plan adaptation when assumptions are violated. We introduce a circuit-breaker mechanism for tool availability and a just-in-time parameter generation strategy that defers input resolution until execution time. The framework is evaluated on multi-node plans involving browser automation, API calls, and structured data extraction across enterprise SaaS environments.

Key contributions

• OODA-loop execution engine with error-driven re-planning for multi-tool coordination
• Parallel and sequential DAG scheduling with dependency-aware node execution
• Circuit-breaker pattern for tool availability with automatic fallback strategies
• Just-in-time parameter generation that adapts inputs based on runtime context