A customer says their AI deployment 'can reason about the problem but never actually does anything'. They are calling a model API directly. Explain what is missing architecturally and what you would build.

1. Draw the boundary: a model is a function from text to text — it has no ability to open a file, call a service, or observe the world. 2. Identify the missing layer: an agent is the software around the model that exposes tools, executes the calls the model requests, and feeds results back. 3. Specify the loop: model proposes a tool call, harness executes it, result is appended to context, model decides the next step, repeat until a termination condition. 4. Define the tool contract: schemas, argument validation, and error results that the model can actually recover from. 5. Add termination and safety controls: step limits, cost limits, and human approval gates for irreversible actions. 6. Decide what stays deterministic in code rather than being delegated to the model — authorization, money movement, and destructive operations. 7. Name the observability you need: a trace of every tool call and result, because failures are loop failures, not model failures.