Distinguish a workflow from an agent
Anthropic distinguishes predefined workflows from systems where models direct their own process and tool use. Retrieval can be part of either. RAG is a technique for supplying reference context, not a guarantee that the output is deterministic or correct.
Begin with the simplest useful baseline
Test a fixed retrieval-and-answer sequence before adding dynamic orchestration. Inspect retrieved material and answer support separately. A missing source passage may require a retrieval fix rather than another planning loop.
Add autonomy only for a defined need
An agent can decide which step to take after observing results, but that also adds operational decisions and possible failures. Define allowed tools, stopping conditions, budgets and human approval for consequential actions. A read/write tool can also be used in a predefined workflow; an agent is not required merely because an API changes state.
Measure both architectures on the same tasks
Record the complete request, tools used, retries, review effort, latency and cost per accepted result. This page reports no universal sub-second response time, accuracy improvement or token multiplier.
| Question | Fixed workflow | Agent loop |
|---|---|---|
| How is the next step chosen? | Predetermined orchestration | Model-directed choices within limits |
| What needs evaluation? | Retrieval and answer support | Those checks plus decisions and tool actions |
| What limits are needed? | Timeouts, permissions and failure handling | Those controls plus explicit loop and spending bounds |
Connect extraction to retrieval quality
Follow the web extraction comparison and vector database evaluation guide to test the data path before adding orchestration.
Frequently asked questions
When is a simple RAG workflow a useful starting point?
When a question can be answered from retrieved reference material through a predictable sequence of steps. Evaluate retrieval relevance, answer support and total response time before adding orchestration.
When should I consider a stateful agent?
When the task needs decisions about intermediate results or a changing sequence of tools. Bound the steps, validate inputs and require appropriate approval before external actions; compare against a simpler baseline.
Sources & further reading
Primary sources checked Oct 17, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Building effective agents ↗Anthropic
Help us keep this useful. Report a factual correction →
