How to Build an AI Agent: A Step-by-Step Guide With a Real Example

Building a demo AI agent takes an afternoon. Building one you can trust with real work takes a process. This guide shows how to build an AI agent the way we do it in client projects, using one running example: an agent that qualifies inbound leads for a B2B startup.

If you want the bigger picture first (architecture, tool choices, and when agents make sense), read our guide to AI agent development.

Step 1: Write the job description

Treat the agent like a new hire. Write down the goal, the inputs, the allowed actions, and what "done" looks like.

Step 2: List the tools

Each tool is a function the model can call, with a clear name, a description, and typed inputs. Keep the list short; every extra tool is another way to go wrong.

Step 3: Build an evaluation set

Before writing prompts, pull 50 past demo requests and record the right outcome for each: the score band and the correct email. This is your test suite. Without it you are tuning by feel.

Step 4: Write the agent loop

At its core every agent is the same loop: send the goal and tool list to the model, run whatever tool it asks for, give it the result, repeat until it says it's done. Frameworks like LangGraph, the OpenAI Agents SDK, and the Claude Agent SDK give you this loop plus retries, state, and tracing. Add three limits from day one:

• A step limit (say, 15 tool calls) so it can't loop forever.
• A cost limit per task, enforced at the model gateway.
• A timeout, after which it escalates to a person.

Step 5: Write the system prompt

The system prompt is the job description from Step 1, made precise: the goal, the rubric, the rules ("never email anyone not in the request"), the tone of the emails, and when to escalate. Include two or three worked examples. Keep business rules in data the agent loads through a tool (like get_icp_rules) so non-engineers can update them.

Step 6: Run the evals and read the traces

Run all 50 examples. For each failure, read the trace (every model call and tool result) and classify why it failed: missing information, wrong tool choice, misread rubric, or bad output format. Fix the most common category first, then re-run. Repeat until results stop improving.

Step 7: Add guardrails

• Approval on irreversible actions. At launch, send_email creates a draft a person approves in one click.
• Output validation. Check the score is a number from 0 to 100 and the email uses an approved template before anything is saved or sent.
• Narrow permissions. The CRM token can only update the lead being processed.
• Audit log. Every action is recorded with the reasoning that led to it.

Step 8: Deploy in shadow mode, then widen

Trigger the agent from your form or CRM webhook. For the first two weeks it proposes actions and a person approves them. Track approval rate, time saved, and cost per lead. When approvals are consistently a formality, remove the approval step for high-confidence cases and keep it for the rest.

Step 9: Monitor and improve

Keep tracing on, alert on failures and cost spikes, and add every new failure to the evaluation set. When a new model is released, run the evals before switching. This is ongoing AI agent development, and it's what keeps the agent useful.

Tools we use to build AI agents

• Models: Claude, GPT, and Gemini, chosen per task, routed through a gateway such as AiKey.
• Orchestration: LangGraph, OpenAI Agents SDK, Claude Agent SDK, or plain code for simple loops.
• Knowledge: Postgres with pgvector, or a managed vector store, plus keyword search.
• Observability: tracing and evaluation tooling, plus cost dashboards per agent.
• Deployment: serverless functions or containers on AWS, GCP, Azure, Cloudflare, or Vercel.

Build it yourself or get help?

If you have engineering time, the steps above will get you a working agent. If you'd rather have it live in a few weeks with evaluation and guardrails done properly, our AI agent development services follow exactly this process. Curious about budget? See our AI agent development cost guide.