Search for a command to run...
Multi-agent AI systems face a fundamental tension: centralized orchestration provides control but lacks adaptivity, while decentralized agent-to-agent coordination enables autonomy but sacrifices traceability and context coherence. Existing approaches - ranging from supervisor-based patterns to the Model Context Protocol (MCP) - address isolated aspects but fail to support iterative, human-guided knowledge work where requirements emerge during execution. We introduce Contractual Task Orchestration (CTO), a hybrid architecture that combines centralized planning with self-organizing execution through a task marketplace. Each task is governed by an explicit contract specifying prerequisites, context requirements, and retrieval scope, enabling agents to dynamically retrieve precisely scoped information rather than maintaining redundant state. A dedicated Framework Validation Agent proactively detects cycles, orphaned tasks, and context inconsistencies, while a versioning mechanism allows runtime adaptation when requirements change - without restarting workflows. CTO supports bidirectional progression: forward (context-to-implementation) for plan-driven work, and reverse (implementation-to-context) for prototype-driven justification. We demonstrate CTO's applicability through a regulatory compliance scenario (maritime vessel inspection system) and position it against related orchestration patterns. Our contribution is threefold: (1) explicit context-scoping as a first-class architectural concern, (2) pull-based coordination with continuous validation, and (3) runtime adaptivity through event-driven versioning. CTO addresses a gap in AI orchestration for collaborative, uncertainty-laden domains such as business engineering, rule-making and public administration, where decisions must remain explainable and auditable under governance and regulatory constraints.