FREQUENTLY ASKED
How is an agent different from a chatbot?
A chatbot primarily generates a response. An agent controls part of a workflow: it selects actions, uses tools, observes results, and continues until an outcome or stop condition. A chat interface can front either system.
Is RAG an agent?
No. Retrieval-augmented generation grounds a model with selected information. It becomes part of an agentic system when a model dynamically chooses whether, how, or repeatedly to retrieve as part of pursuing a goal.
When should I not use an agent?
When the path is stable, rules are clear, error costs are high, or a deterministic workflow meets the requirement. Agents trade cost, latency, and predictability for flexibility.
When do multiple agents make sense?
When distinct domains need separate context, tools, ownership, or evaluation—and a single agent is demonstrably overloaded. Multi-agent design adds coordination failures, latency, cost, and a larger security surface.
Can guardrails eliminate hallucinations?
No. Guardrails reduce specific risks. Reliability comes from constraining the task, grounding inputs, deterministic validation, good tools, evaluation, human oversight, and safe failure behavior.
How do I measure agent ROI?
Compare against the current workflow: successful outcomes, quality, cycle time, labor saved, total cost, escalation rate, and downstream errors. Include review and incident costs—not only token spend.
Should I fine-tune the model for my agent?
Usually not first. Begin with clear instructions, better context, high-quality tool descriptions, retrieval, and eval-driven iteration. Fine-tuning becomes useful when you have a stable task, representative examples, a measurable baseline, and a repeated behavioral gap that prompting cannot economically solve.
Do I need an agent framework?
No. A small loop with structured model output and well-defined tools is often the best first implementation. Adopt a framework when its tracing, state, orchestration, deployment, or interoperability primitives remove demonstrated engineering work—not because an agent requires one.
What is the difference between MCP and A2A?
MCP standardizes how AI applications connect to tools, resources, and prompts. A2A standardizes how independent agents discover capabilities, exchange messages, manage tasks, and return artifacts. They address different connection layers and can be used together.
Why is indirect prompt injection dangerous?
An agent may encounter malicious instructions inside webpages, documents, emails, or tool results. If it treats that untrusted data as authoritative guidance, it can leak information or misuse tools. Isolate trusted instructions, label provenance, constrain capabilities, and authorize actions outside the model.
What makes an agent production-ready?
A measurable task contract, versioned instructions and tools, scoped identity, adversarial evaluations, end-to-end traces, budgets, approval policy, safe retries, incident ownership, staged rollout, and a tested fallback. A compelling demo is only evidence of possibility.
Can deterministic workflows and agents coexist?
Yes—and they usually should. Keep policy, permissions, validation, arithmetic, and known process steps deterministic. Use model judgment for ambiguous interpretation, planning, and selection. The strongest systems combine both deliberately.
