Choosing an AI solution architecture
Start with the task and measurement
Choose the simplest solution that meets the requirements for quality, safety, response time and cost. Some tasks need only rules or a single model call. Multiple agents add overhead, failure points and review requirements.
A frontier model is a model at the technological frontier; the term alone does not identify the best model for a particular task. A smaller general-purpose model may work without custom training. Running a model yourself is not necessarily cheaper than an API: include hardware, utilisation, administration, updates and security.
Four options
| Pattern | When to consider it | What to verify |
|---|---|---|
| Rules and conventional automation | The steps and desired outcome can be defined precisely | Exception coverage and rule maintenance |
| One model with the necessary context | The task fits into one well-defined step | Quality on your own examples and access to sources |
| A fixed workflow or routing | Different inputs need different tools or models | Routing mistakes, handoffs and the whole process cost |
| An orchestrator and worker agents | Subtasks depend on the request and cannot be specified well in advance | Agent permissions, coordination, limits and the quality of the final synthesis |
An orchestrator can allocate work and assemble a result. A worker agent can use a model, search or a conventional program. Assigning a large model to reasoning and a smaller one to execution is an option to test, not a general rule.
Example: drafting a response to a customer complaint
- Retrieve data you are authorised to access and establish the request type.
- Find the current terms and the necessary case history.
- Draft a response with references to the supporting information.
- Check factual details and the authority to offer a remedy.
- Have a person approve risky or ambiguous cases before sending.
First test this process with one model. Add routing to a smaller model or multiple agents if measurement demonstrates value. There is no generally established “20% of cost / 80% of value” ratio for this example.
Pilot criteria
Define representative tasks, failure cases and acceptable quality in advance. Compare alternatives on the same inputs. Track the whole case cost, including corrections, time to a usable result, incorrect actions and human interventions. Repeat the evaluation after a model change.
The patterns follow the general distinction between workflows and agents in Anthropic’s technical article; your own evaluation must determine the specific choice.