METANOWBegin with the model and its context
A model processes an input and generates an output. The surrounding application determines which instructions, documents and conversation history it receives. Context is limited, so selecting relevant information matters.
Add tools and working state
Tools provide defined operations, such as searching an approved repository. State records what has happened during the task. Persistent memory is optional and requires an explicit purpose, access rules and retention decisions; it is not an automatic property of every agent.
Coordinate the steps
An orchestrator controls routing, tool calls, retries, hand-offs and completion. Some paths are predetermined; others allow model-led choices. The application must check tool inputs and results rather than trusting a generated request.
Make the result reviewable
Keep evidence, tool outcomes, workflow versions and reviewer decisions. Evaluate task success and failure modes, including unsupported answers and unauthorised actions. Human approval should be enforced in the application where a decision requires it.
Sources & further reading
Reviewed 19 October 2026. This is a technology learning session; examples describe possible designs, not a claim that a particular integration is active.