Build an agent

You can write the loop yourself with a provider's SDK. Send the chat, run the tool the model asks for, add the result, and repeat. Anthropic's guide suggests you start this way. A framework can hide the real prompts and replies, so bugs are harder to find.

An agent framework helps when you need what it already solved. Examples are retries, saved state, human approval, tracing and hand-offs between agents. As of October 2026, options include the Claude Agent SDK, the OpenAI Agents SDK and LangGraph. Others are Google's Agent Development Kit, Microsoft Agent Framework, CrewAI, the Vercel AI SDK and Pydantic AI. They are in no order, and they change fast.

Choose by your team's language and the model providers you need. Also think about how much control you want. Whatever you pick, learn what it sends to the model.

• Few tools. More tools do not always help. OpenAI's docs loosely suggest fewer than 20 at a time. Tools that overlap cause more mistakes than many tools. Join steps the model always does together into one tool.
• Clear names. Say what the tool does and when not to use it. Add a system name to similar tools, like orders_search and kb_search.
• Inputs that are hard to misuse. Name it user_id, not user. Give a fixed list of choices, not free text.
• Short results. Thousands of rows fill the context window and the bill. Filter first.
• Useful errors. "PIN code must be 6 digits. You sent 56004." gets a fix. "Error 422" gets the same failing call again.

A run can stop halfway. The server restarts, or the run waits two hours for an approval. It should go on from a save point. It must not start again and refund twice. This is durable execution.

Keep long-term memory per signed-in customer. Then nothing about Sam shows up in Maya's chat. Treat saved memory as untrusted. A note saved from an email can carry orders into later runs (prompt injection).

Pause for a person before anything you can't undo, anything that spends money, or anything that sends data out. Show the exact action, like a ₹1,499 refund to Maya's card for order 1042. Ask the right person (human-in-the-loop approvals). Too much power for the task is excessive agency.

Some old systems have no API. Computer use lets a model work a screen. Your code sends a screenshot. The model replies with an action, like a click. Your code does it and sends a new screenshot. As of October 2026, Anthropic, OpenAI and Google all offer this.

It is slow, because every step sends an image. It breaks easily, for example on a pop-up or a redesign. It is risky, because it acts with the browser's logins. So run it in a closed-off box with few logins and a list of allowed sites. Cap steps and cost. Ask a person before it buys or sends anything. If an API exists, use the API.