What is an AI agent?
A normal model call is like asking one question and getting one answer. An agent is more like asking someone to work through a task. It can look at context, choose a tool, check what happened, and decide the next step.
AI agents FAQ
Plain answers for teams learning how agents work.
A practical guide to the ideas behind modern AI agents: context, tools, MCP, memory, RAG, approvals, guardrails, and proof.
Start here
AI agents sound complicated because people describe the machinery first. The simple version is this: an agent is software that can use a model, remember context, call tools, and keep working toward a goal.
When do I need an agent instead of one model call?
Use a direct model call for a one-shot job, like turning a transcript into a summary. Use an agent when the work has steps: search, compare, call a tool, wait for a result, revise, and keep going.
Are agents fully autonomous?
Autonomy is a spectrum. Some agents only choose between a few safe paths. Others remember work, call tools, retry failures, and plan across many steps. The more autonomy you give, the more control and proof you need around it.
Context, memory, and retrieval
The biggest practical lesson is that more context is not automatically better. The agent needs the right context at the right time.
What is a context window?
The context window is the amount of text a model can read at once. Bigger windows are useful, but they are not magic. If you pack the window with stale files, old notes, and unrelated tool output, the agent can still miss the thing that matters.
What is RAG?
RAG means retrieval-augmented generation. You break documents into chunks, turn them into searchable embeddings, retrieve the most relevant pieces for a question, and give those pieces to the model before it answers.
What is agent memory?
Memory is how an agent carries useful facts forward. It might remember stable preferences, retrieve similar past information, or compress a long session into short observations so the next step starts with the right background.
How should teams manage context cost?
First make the workflow work. Then make it reliable. Then make it cheaper and faster. Once the task is understood, teams can prune old messages, cache stable prompts, route simpler steps to smaller models, and retrieve only the context that matters.
Tools, MCP, and workflows
A tool is where an agent crosses from words into action. That is why tool design matters so much.
What are tools in an agent system?
Tools are functions an agent can call: search a database, fetch a ticket, write a file, send a message, run a check, or call an API. Clear tools make agents easier to guide because the model knows what each tool does and when to use it.
What is MCP?
MCP is a standard way for agents and tools to connect. Think of it as a common adapter: a tool can expose its capabilities once, and MCP-compatible agents can discover and use those tools in a predictable way.
What is a workflow graph?
A workflow graph breaks the job into steps. Some steps run one after another. Some branch into parallel checks. Some merge results. Some pause until a human or external system responds.
Why do workflows need suspend and resume?
If an agent needs a human approval or a slow external result, it should not keep a process running forever. A better workflow saves its state, waits, and resumes from the same point when the answer arrives.
Safety, proof, and rollout
The moment an agent can change real systems, the question becomes: who allowed it, what did it see, what did it do, and can we review the trail later?
What are guardrails?
Guardrails are checks around the agent. They can filter unsafe inputs, block off-topic requests, protect private data, and stop the agent from following malicious instructions hidden in documents, tickets, or web pages.
When should a human approve an action?
Ask for approval before high-risk or hard-to-undo actions: deleting data, writing to production, sending external messages, using credentials, changing configuration, or running code with broad access.
What should be logged?
Log the agent, user, context used, tool requested, policy decision, approver, reason, and result. That turns an agent run from a blurry conversation into a reviewable record.
What is the best way to ship agents?
Start narrow. Give the agent a clear job, clear tools, and clear boundaries. Watch where it fails. Add approvals where the stakes rise. Then use evals and audit trails to decide when it is ready for more autonomy.