Design business applications, shared knowledge, verification, and approvals together. Specialist agents use the knowledge and tools their work requires and connect results to the next action.
Instead of concentrating every domain, tool, and permission in one giant agent, connect specialists through a coordinator where needed. Let business requirements and evaluations determine agent boundaries. A single agent or explicit workflow may be sufficient for simpler work.
Profitability analysis and supply planning need different expertise, calculations, and change permissions. These differences can justify separate agents. Question types such as diagnosis, prediction, and simulation do not each require their own agent.
Context Boundary = Responsibility / Permission Boundary is a principle for aligning scope, not a security guarantee. Data services and tools must enforce actual permissions. A coordinator does not automatically inherit every agent’s authority.
Manage state, tools, permissions, and evaluations per agent while reusing shared model services and runtime infrastructure.
Use validated tools for calculations and rules, and agents for interpretation and exploration. Automate pre-authorized actions and obtain approval where policy requires it. Feed outcomes and recurring exceptions into reviewed knowledge and operating criteria.