Eight layers every agentic product needs, independent of the specific model or stack. I wanted this section to explain not only what changed, but why the previous approach stopped being enough once the product met real operating conditions.
Input
Reasoning
Tools
Execution
Feedback
Persistence
Control
Commercial outcome
Input
What forms of input does the product accept (text, image, URL, code, video)? I would answer this with the actual current system behaviour, not the intended design on a diagram.
How is ambiguous or malicious input handled before it reaches the model? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Reasoning
Is this a single call or a multi-turn loop? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Is the orchestration coupled to one model provider? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Tools
What can the model actually do, concretely (files, commands, network)? I would answer this with the actual current system behaviour, not the intended design on a diagram.
What is the policy boundary around each tool? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Execution
Where does generated or agent-driven work actually run? I would answer this with the actual current system behaviour, not the intended design on a diagram.
What isolation and resource limits apply to that execution? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Feedback
How does the system know an action failed? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Is failure fed back to the agent in a structured, actionable form? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Persistence
What is the durable object the user expects to keep? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Does history, diffing, and rollback exist for it? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Control
Where can a human inspect, pause, or reverse what the agent did? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Are customer-facing actions gated by human approval where it matters? I would answer this with the actual current system behaviour, not the intended design on a diagram.
Commercial outcome
What business process does this feed into once generation is done? I would answer this with the actual current system behaviour, not the intended design on a diagram.
How is usage, cost, and value measured per customer? I would answer this with the actual current system behaviour, not the intended design on a diagram.