Before designing automation, decide whether you need a fixed workflow, a model call or an agent that chooses intermediate steps. This article explains that architecture decision. It is separate from our step-by-step small-business enquiry pilot.
The distinction
Anthropic’s December 2024 engineering guide distinguishes predefined workflows from agents whose models direct their process and tool use. It describes simplicity, cost and error trade-offs; it does not establish that every task needs an autonomous agent.
| Task characteristic | Starting design to consider | Question before adding complexity |
|---|---|---|
| Known fields and fixed steps | Deterministic workflow with validation | Can ordinary rules handle the exceptions? |
| Unstructured text with one bounded output | A model call followed by validation | Can you check the result and handle missing evidence? |
| Steps genuinely depend on intermediate findings | Agent with narrowly scoped tools | Who controls its actions, stopping conditions and failure recovery? |
This table is editorial design guidance, not a benchmark or a product availability claim.
Read access and write access are different risks
Retrieving an approved document differs from sending a message, changing a record or spending money. Specify each permitted action in the system that grants access. A prompt saying “be careful” is not a substitute for an enforced permission boundary.
Keep an unresolved request visible for review. Set stopping conditions for tool failure, conflicting evidence or an action outside the allowed scope. Do not let a repeated attempt silently create a second transaction.
Do not assume feedback means learning
The 2023 Reflexion research paper investigates agents using verbal feedback on specified tasks. It does not establish universal hallucination reduction or that every commercial agent updates its behaviour when a user corrects it. Inspect how a chosen system stores feedback and changes instructions.
Three questions for a design review
- What evidence tells us that the output is acceptable?
- What happens when a tool returns nothing, contradicts another source or times out?
- Which action must wait for a person, and where is that enforced?
Keep deployment details distinct from the architecture. Local model processing does not establish that all logs, account services or connected tools remain local.
For a concrete FAQ-based draft-and-review workflow with acceptance tests and an illustrative cost worksheet, use our small-business agent pilot guide. For assessing whether a tool is worth buying, use the AI tool evaluation guide.
