3.1.3. Prompt and Response Agents
Prompt and response agents are the simplest agent type — they receive a user input, process it using AI, and return a response. No multi-step workflows, no background monitoring, no persistent state between interactions. Each interaction is stateless and self-contained.
Despite their simplicity, prompt-and-response agents serve critical business functions. They're the right choice when the interaction pattern is straightforward: a user asks a question, the agent answers. FAQ bots, knowledge retrieval agents, and simple data lookup agents all fit this pattern.
Design Characteristics:
| Characteristic | Prompt-and-Response Behavior |
|---|---|
| Initiation | User sends a message |
| Duration | Single request-response cycle |
| State | Stateless (or minimal conversation context) |
| Decision-making | Response generation only |
| Human involvement | Every interaction is user-initiated |
| Completion | Response delivered = interaction complete |
When to Use Prompt-and-Response:
- The user needs a quick answer, not a workflow execution
- No downstream actions are required (no record updates, no approvals, no notifications)
- The interaction is self-contained — previous interactions don't inform the current one
- Response speed is critical — minimal orchestration overhead
When NOT to Use:
- The user's request requires multi-step processing ("process this invoice" = task agent)
- The interaction needs to persist state across messages within a conversation
- Actions need to be taken on external systems based on the response
- The scenario requires proactive behavior (= autonomous agent)
Design Considerations:
Prompt-and-response agents still require careful design:
- Knowledge sources must be configured to ground responses in accurate data
- System prompts define the agent's persona, boundaries, and response format
- Fallback behavior determines what happens when the agent can't answer (escalate to human, suggest related topics, or acknowledge the limitation)
- Output quality depends on prompt engineering — clear instructions, constraints, and formatting guidelines
Optimizing for speed: use a small language model (SLM) instead of a large one when it meets the quality bar, ground on pre-indexed knowledge sources rather than real-time API calls so retrieval is an index lookup, and keep the system prompt tightly scoped. Chaining several agents, or dropping grounding to rely on model memory, works against latency or accuracy.
When a "single question" needs multi-step reasoning: a question that combines several regulations is still prompt-and-response. Keep the architecture and add step-by-step reasoning instructions to the system prompt and retrieve more passages (a higher top-K); a task agent is warranted only when real actions or workflow steps are involved. Statelessness is the other limit — if users need follow-ups that remember earlier turns, redesign for multi-turn context rather than upgrading the model.
Exam Trap: The exam may describe a scenario that sounds like it needs a task agent ("help the user find the right product") but is actually a prompt-and-response pattern — the agent recommends a product based on the user's description, and the user makes the decision. If no downstream action is automated, it's prompt-and-response.
Reflection Question: A company wants an agent that answers employee questions about company policies by referencing an internal knowledge base. No actions need to be taken — just answers. But employees also want to ask follow-up questions in the same conversation. Does this change the agent type? Why or why not?