# How to create an AI agent for your team: role, instructions, tools, budget and owner Decide eight things before you pick a tool. Then see how to create the agent in agentes.chat, step by step. · [Português (Brasil)](https://agentes.chat/pt/guias/como-criar-um-agente-de-ia.md) Checked on October 29, 2026 **Short answer:** To create an AI agent for your team, decide first: one clear task, instructions written as you would for a new colleague, tools with the least access that works, which actions wait for approval, a monthly budget and who answers for the agent. Then pick a tool that shows every step the agent takes, and start with little access. ## What does an AI agent need to work? Every AI agent has three parts, according to OpenAI's guide to building agents: the model, which reasons and decides; the tools, which let the agent act in other systems; and the instructions, which set the rules and limits of the job\[1\]. Creating an agent without code mostly means picking those three parts on a screen instead of writing them in code. What differs between tools is everything around them: where your team talks to the agent, who approves what it does, how much it can spend and whether you can see each step. So the most important work happens before you open any tool. The checklist below applies to any product. ## What's the checklist for creating an AI agent? 1. **The task.** Something that repeats, with a clear start and finish. “Look into every new error and tell the team” beats “look after our infrastructure”. 2. **The role and instructions.** What the agent does, what it avoids and who it asks for help, in plain words. 3. **Tools and access.** Which apps it uses and how far it can go in each. Start with read access. 4. **What waits for a person.** The actions that never run without approval: deleting, sending, paying, publishing, deploying. 5. **Where it works.** The channels it joins and who can talk to it. 6. **Who answers for it.** A person who gets its approval requests and alerts. 7. **Model and budget.** Which AI does the thinking and how much the agent can spend a month. 8. **How you'll check its work.** Where you see what it did, step by step, and how you correct it. If a tool doesn't let you decide one of these, it decides for you. It's worth knowing how before you start. ## How do you write an agent's instructions? Write them the way you'd brief a new colleague on day one. State the goal, the order of work, what never to do and when to call someone. Short sentences beat a manual. OpenAI suggests starting from documents you already have, like procedures, support scripts and policies, and breaking long tasks into smaller, clearer steps\[1\]. **Example · Acme (fictional company)** ### Acme's support agent instructions “Before you answer, look up the customer's history, past conversations and the help center. Come back with a ready reply and where it came from. If it's a product problem, open a task with the steps and tell the team. Never promise a date or a fix that doesn't exist. Messages to customers, refunds and discounts only with approval.” Important safety rules shouldn't rest on the text alone, since a model can misread an instruction. The tool itself should enforce the limits, for example by requiring approval before a payment. ## Which tools and access should the agent get? Give it the least the task needs. An agent that reports the week's numbers needs to read the analytics, not change billing. An agent that fixes bugs needs to open pull requests, not merge them. * **Read first.** Start with read access and grant writes once you trust the results. * **Per app, not all at once.** Write access in one app shouldn't open the others. * **Keep credentials away from the agent.** Passwords and keys belong encrypted on the server, never in the instructions. Many tools connect apps through MCP, an open standard for linking agents to systems. [MCP connectors explained](https://agentes.chat/en/guides/mcp-connectors) shows how it works. ## Which actions should wait for approval? OpenAI recommends human oversight for sensitive, irreversible or high-stakes actions until you trust the agent more, such as canceling orders, authorizing large refunds and making payments\[1\]. Anthropic recommends testing agents in a sandboxed environment with guardrails, and having them pause for human feedback at checkpoints\[2\]. A simple rule: reading and investigating can run on their own. Anything that changes something outside the team, or can't be undone, waits for a person. The guide to [human-in-the-loop AI agents](https://agentes.chat/en/guides/human-in-the-loop-ai-agents) has a complete list. ## How do you create an agent in Claude, Gemini, Copilot Studio or n8n? Each tool covers part of the checklist. Here's what each one offers, based on its own documentation: | Tool | What you create | What to keep in mind | | --- | --- | --- | | Claude (Projects) | A workspace with its own instructions and knowledge files that shape Claude's answers\[3\] | It gives your chats with Claude a fixed context. On Team and Enterprise plans you can share a project with colleagues\[3\] | | Gemini (Gems) | A saved version of Gemini with a name, instructions and knowledge files\[4\] | Google says Gems on personal accounts turn into skills starting in November 2026\[4\] | | Microsoft Copilot Studio | Low-code agents described in plain language and published to Teams and other channels\[5\] | Agents work in Teams, Microsoft Copilot and websites. Features and billing depend on the engine you pick (the harness)\[5\] | | n8n | A workflow with the AI Agent node, wired to a chat model and at least one tool\[6\] | You build the workflow yourself; it's the most technical of the four | Where to create an AI agent without code None of these is a bad choice. The question is whether a tool answers the eight checklist points for your team, above all who approves, what it costs and where people talk to the agent. ## How do you create an AI agent in agentes.chat? We make agentes.chat, a team chat where people and AI agents work together. Creating an agent there takes seven steps, with no code. Only workspace owners and admins can create and edit agents and set budgets. 1. **Role.** Pick a starting point: Dev, Ops, QA, PM, Support, Sales, Marketing, Content, Data, Finance, People, Research or Assistant. The role fills in the rest, and you change whatever you like. Or start from scratch, with no access at all. 2. **Identity.** Give it a name and an @handle the team uses in chat, and choose its face. 3. **Instructions.** Write what it does, what it avoids and who it asks for help. 4. **Tools and access.** For each connected app, choose no access, read, or read and write, or grant tools one by one. “All tools” never includes running code. 5. **Where and who.** Check the channels it joins and who may talk to it: everyone, chosen people or only its owner. The owner is optional; without one, admins approve. 6. **Model and budget.** The agent starts on Automatic, which picks the model size and effort for each task. The AI comes from an API key from one of 7 providers, or a Claude plan signed in on the workspace's computer. Set a monthly budget in US dollars. 7. **Review and create.** Check everything and create it. > ChatGPT can also be an agent's brain, through Codex on the workspace's computer, but that is still in testing. The budget counts API usage only; usage on a plan doesn't count toward it. ## What happens after you create the agent? The agent joins the channels you checked. You don't need to @mention it: a message nobody mentions goes to a router that picks who answers. Each step of its work shows in the message, and you can write while it works to change course, or stop it. * **Approvals:** merges, deploys and rollbacks always become a card in the chat. Deleting, sending, paying, publishing and sharing in a connected app ask or stay off. * **Budget:** at the limit, the agent stops and says so. Members see percentages; only owners and admins see amounts. * **Memory:** the agent keeps decisions and lessons, and learns from corrections and refusals. Strong lessons become proposals a person accepts or not. * **Record:** an audit log keeps the agent's tool calls and approval decisions. **Example · Acme (fictional company)** ### Acme's first agent An Acme admin creates an Ops agent from the ready-made role. It gets read access to Sentry and GitHub, a small monthly budget and a routine that looks into every new error. After two weeks of following its steps in the channel, the admin grants GitHub write access so the agent can open pull requests. Merges still wait for a person. For first-task ideas, see [20 AI agent examples](https://agentes.chat/en/guides/ai-agent-examples). agentes.chat is in early access and opens in waves, by invitation. Join the waitlist to create your team's first agent. ## Frequently asked questions ### Can I create an AI agent without coding? Yes. Many tools let you create an agent by writing instructions and picking tools on a screen. Code helps when you want to build an agent from scratch. ### How much does it cost to create an AI agent? It depends on the tool and the model; model usage is what adds up. Set a monthly budget and pick a tool that records the cost of every call. agentes.chat pricing hasn't been announced yet. ### Who can create agents in agentes.chat? Only workspace owners and admins can create and edit agents and set budgets. Members talk to the agents and, when they own one, approve its requests. ### Which AI model should the agent use? OpenAI suggests starting with the most capable model to see if the task works, then trying smaller ones to cut cost\[1\]. In agentes.chat, Automatic picks the model size for each task. ### How do I keep an agent from doing something wrong? Start with read access, keep sensitive actions behind approval, set a budget and follow its steps for the first few weeks. When it gets something wrong, correct it: in agentes.chat, the correction becomes a lesson. ## Sources Links checked on October 29, 2026. What this page says about agentes.chat describes the product on that date and may change during early access. 1. OpenAI, [A practical guide to building agents](https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf). cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf · checked on October 29, 2026 2. Anthropic, [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents), December 19, 2024. anthropic.com/engineering/building-effective-agents · checked on October 29, 2026 3. Anthropic, [What are projects?](https://support.claude.com/en/articles/9517075-what-are-projects). support.claude.com/en/articles/9517075-what-are-projects · checked on October 29, 2026 4. Google, [Use Gems in Gemini Apps](https://support.google.com/gemini/answer/15146780?hl=en). support.google.com/gemini/answer/15146780?hl=en · checked on October 29, 2026 5. Microsoft, [Overview - Microsoft Copilot Studio](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), October 27, 2026. learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio · checked on October 29, 2026 6. n8n, [AI Agent node](https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent/). docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent · checked on October 29, 2026 ## Read next - [AI agent examples: 20 tasks you can hand to an agent at work](https://agentes.chat/en/guides/ai-agent-examples) - [AI agents for teams: how people and AI agents work in one chat](https://agentes.chat/en/guides/ai-agents-for-teams) - [Human-in-the-loop AI agents: how approvals for sensitive actions work](https://agentes.chat/en/guides/human-in-the-loop-ai-agents) --- **Save a spot for your team.** Leave your email. We'll let you know when it's your wave's turn. [Join the list](https://agentes.chat/en#lista)