How to Set Up an AI Agent

“AI agent” has become a catch-all term, and the setup steps depend entirely on which kind you actually need. A Zapier workflow that summarises your inbox is an agent. Claude Code refactoring your codebase is an agent. A LangGraph state machine orchestrating five tool calls is an agent. The work you do to stand them up is not the same.

This guide is a decision tree. In ten minutes you’ll know which of four paths fits your problem, roughly what it costs, and what to read next. Written April 2026, against the current generation of tools (Claude Opus 4.7, GPT-5 class models, OpenAI AgentKit, Claude Agent SDK 0.1.48).

Step 1 — Name the job before naming the tool

Write one sentence: “I want an agent that takes X and produces Y without me watching.” If you can’t finish that sentence, you don’t need an agent yet — you need a prompt. Try our prompt improver and the prompt generator first. Agents are for multi-step tasks with tool use, not single answers.

Step 2 — Pick the right lane

There are four real options in 2026. Pick the highest-level one that can do the job. Going lower buys flexibility but costs weeks of your time.

Lane A — Use a hosted agent product (1 hour)

For most knowledge-work tasks, the answer is ChatGPT agent mode, Claude’s computer use, or a vertical product like Sierra, Intercom Fin, or Zapier Agents. Zero code, one configuration screen, live in under an hour. See our ChatGPT agent mode guide — it covers 70% of “please do this for me” use cases.

Lane B — Use a no-code platform (1 day)

If the task involves glueing apps together — Slack to Gmail to a sheet — use Zapier Agents, n8n, or OpenAI’s Agent Builder. You draw the flow, pick tools from a menu, and the agent runs on their infrastructure. Recurring cost, but you don’t own an ops burden.

Lane C — Use an agent SDK (1 week)

If you’re a developer and you want your code orchestrating the agent, use the OpenAI Agents SDK (pip install openai-agents) or the Claude Agent SDK(pip install claude-agent-sdk, version 0.1.48 as of March 2026). Both give you Agents, tools, handoffs, guardrails — and you host them yourself. See our OpenAI SDK guide and Claude SDK guide.

Lane D — Use a framework (2+ weeks)

For multi-agent, long-running, or stateful workflows, use LangGraph (stateful graphs), CrewAI (role-based teams), AutoGen (multi-agent conversations), or MetaGPT (software teams). These give the most control and the most rope to hang yourself with. Only pick this lane if you’ve already outgrown Lane C.

Step 3 — Decide where it runs

Three choices. Hosted: the vendor runs it (ChatGPT, Claude, Zapier). Easiest, least flexible, per-seat or per-task pricing. Your laptop: run the SDK locally against an API key. Fastest to iterate, terrible for production. Your server: a small Python service on Fly, Render, or a VPS — the practical choice once the agent actually does useful work.

Step 4 — Wire it to the tools it needs

Almost every useful agent needs to reach outside its model: read a file, call an API, query a database, send a message. In 2026, the standard for this is MCP (Model Context Protocol). See our MCP setup guide — install a server, allow-list the tools, and the agent can use them the same way across Claude, OpenAI, and most frameworks. MCP is the single biggest quality-of-life improvement in agent engineering this year.

Step 5 — Start with read-only, add writes slowly

The single most common production failure is giving an agent write access too early. Begin with read-only tools: let it query, summarise, and suggest. Once you trust its judgment over ten or twenty runs, add one write tool at a time (send message → create draft → send draft). Every Claude and OpenAI SDK supports explicit allow-listing; use it.

Step 6 — Budget for tokens and cap them

An agent that loops can burn $50 in an afternoon. Before you ship, set a per-run token ceiling and a per-day spend cap at the provider level. Run your prompts through our AI token counter so you know the baseline cost before you loop it 100 times.

Step 7 — Measure and iterate

Log every run: inputs, tool calls, outputs, token spend, user feedback. The best agents in production look boring — a narrow task, a tight tool set, 95%+ success, a human reviewing the 5%. The worst look impressive in demos and explode the first time they meet a real user.

The honest shortlist

If you asked me to pick just three places to start today: ChatGPT agent mode for personal research and browsing tasks, Claude Code for anything involving your codebase, and Zapier Agents for cross-app automation. Everything else is for when those three stop being enough.