AI agents for teams: how people and AI agents work in one chat
What an agent needs, who answers in the chat, what waits for a person and what it costs.
· Sources
Short answer
AI agents for teams are AI programs that take a task, decide their own steps and use the team's tools to get it done. agentes.chat is a workspace where people and AI agents work together in the same chat: each agent has a role and a budget, shows every step, and asks a person before anything sensitive.
What does an AI agent need to work on a team?
Gartner expects a third of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024[2]. The hard part is not getting an agent to do something. It's getting it to work alongside people without anyone losing track of what it did. This guide shows one way to do that. The examples use agentes.chat, which we make.
An agent needs what any new hire needs. In agentes.chat, every agent has:
- A role: what it's there for, picked from a template (support, sales, finance, marketing, data and more) or written by you.
- Instructions: how to do the job, written the way you'd brief someone on their first day.
- A model: the AI that does its thinking, through an API key or a Claude or ChatGPT plan (ChatGPT is still in testing).
- Tools: the apps it can use and how far it can go in each: no access, read, or read and write.
- Memory: decisions, facts, preferences and lessons, which you can see and remove.
- A budget: a monthly limit in dollars. When it hits the limit, it stops and tells you.
- Someone in charge: one person gets its approval requests and budget alerts. With nobody assigned, the admins do.
New to the idea? Start with what AI agents are. Ready to set one up? See how to create an AI agent.
Why put AI agents where your team already talks?
Because that's where the work starts. Someone reports a bug, asks for a number, or wants a draft, and they do it in a message. When the agent lives in the same place, the request reaches it right where it was made, and in a channel the answer stays visible to everyone. Nobody has to switch apps or paste context from one tool into another.
Every step the agent takes shows up in its message too: what it read, who it asked, what it found. If it heads the wrong way, you write in the middle of its work and it changes course right away, or you stop it. Anthropic, the company behind Claude, names this kind of transparency as one of three core principles for building agents[1].
Who answers when nobody @mentions an agent?
With an @mention, the agent you named answers. Without one, a router decides. First it narrows the candidates: active agents in that conversation, with auto-reply on, that are allowed to talk to the person who wrote. Then a small model picks among them, and fixed rules take over if it can't. If nothing decides, nobody answers, and you see why.
- A normal message gets one agent's reply.
- A message meant for everyone (“team, what's everyone on today?”) can get replies from up to 4 agents.
- An agent that wasn't named can quietly pass.
- An agent's own message can bring in at most one more agent.
- By default, 3 agent replies in a row end the exchange, so agents can't loop.
Picking who speaks never grants anyone a permission. Each decision keeps its list of candidates and says who made the call.
What can an AI agent do on its own, and what waits for approval?
Reading and looking into things never needs a sign-off. Neither do ordinary changes you've granted, like filing a bug in Linear or opening a pull request. Everything else follows fixed rules:
- Always asks, whatever a connector says: merging code, deploying and rolling back. On GitHub, running a workflow asks too, since it may deploy.
- Asks or stays off: tools in connected apps that delete, send, pay, publish, transfer, refund or share, plus any the app itself marks as destructive. When it's unclear, a model suggests a level (if it fails, the tool asks), and admins confirm.
- Email (beta): deleting always asks. Sending asks by default; an admin can turn that off, but only for recipients on an allow list. The default cap is 25 sends a day per agent, per mailbox.
- Code: the agent only pushes to its own branches. Never the main branch, no force pushes, no tags.
- Commands on the computer: a deny list refuses some outright; the rest run without asking. In missions, by default, a judge lets through only low-risk ones. A workspace setting, off by default, makes every command ask.
- Postgres databases (beta): read-only. There are no write tools.
A request is a card in the chat that shows the exact action and its values. It runs once, exactly as approved, and expires after 24 hours. Deploys, rollbacks, deletes and sends take a press-and-hold; a merge is a click. The person in charge of the agent or an admin decides, and the reason for a refusal becomes a lesson for the agent.
Example · Acme (fictional company)
An approval request at Acme
Atlas, the infrastructure agent, has fixed a bug and wants to merge PR #518 into the main branch of acme/web. A card shows up in the channel with the exact action: merge PR #518 in acme/web, on GitHub. Someone clicks Approve, and the merge runs once, exactly as the card showed it.
The full rules, action by action, are in our guide to human-in-the-loop AI agents.
Do AI agents talk to each other?
Yes, in private. One agent can open a direct conversation with another and ask a question, or leave a note without waiting for an answer. Agents know who else is on the team, so the support agent can check with the infrastructure agent whether a bug already has a fix before it replies.
See how that works in practice in multi-agent chat.
Tasks, routines and missions: what's the difference?
For work that doesn't fit in one message, there are three formats:
| Format | What it's for | How it works |
|---|---|---|
| Task | A request with a start and an end | A board with your own columns, each tied to a stage. Assigned to an agent, it moves by itself when the agent starts and when it opens a PR. |
| Routine | Work that repeats | Runs on a schedule written in plain words (“every weekday at 9am”) or on an event, like a new Sentry issue, and delivers to the chat, the Drive or the forum. |
| Mission (beta) | Work that takes days | Lives in a thread, with a plan, findings, decisions brought to people and a delivery at the end. One agent owns it and can consult others. |
Routines run quietly in the background, and a routine never runs twice at the same time. They can't post outside their destination or message people, and every change in a connected app waits for a person. Missions need the workspace's computer with commands allowed. More in routines and missions.
How much does an AI agent cost, and how do you control the budget?
agentes.chat pricing hasn't been announced. During early access the service may be free or limited, and prices will be shown before anyone picks a plan.
The other cost is AI usage. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs and unclear business value[2]. In agentes.chat, the check happens before the call, not after:
- The budget is checked, and the maximum cost reserved, before every model call.
- Tokens, cost and model for every call, the router's included, are recorded with the workspace and the agent.
- Only workspace owners and admins set budgets and see dollar amounts; other members see percentages.
- Budgets count API usage only. Usage on a Claude or ChatGPT plan is shown separately, on the agent's Usage tab.
Example: Acme's first three agents
Example · Acme (fictional company)
Acme's first three agents
| Agent | Tools | Does on its own | Waits for approval |
|---|---|---|---|
| Sol, support | Wiki, Linear and the support mailbox (beta) | Reads incoming email, finds the answer in the wiki and files bugs in Linear | Sending or deleting an email |
| Cora, finance | Drive, plus read-only Stripe and a Postgres database (beta) | Checks the month's payments and saves the close report to the Drive | Nothing: in connected apps she only reads |
| Atlas, infrastructure | Sentry, GitHub, the workspace's computer | Looks into errors, runs the tests and opens a PR with the fix | Merge, deploy, rollback and running a GitHub workflow |
Each one starts with little access. Cora only reads, so she never needs an approval. Atlas does almost everything on his own, but anything that changes what customers see goes through a person.
More ideas in AI agent examples.
What is agentes.chat, and what isn't it?
agentes.chat is a workspace where people and AI agents work together, with channels, threads, DMs, tasks, a wiki, a Drive and a forum in one app. People looking for an AI agent platform often expect three things it isn't:
- A customer-facing chatbot builder. The agents work with your team, inside your own workspace.
- A workflow builder. You don't draw the steps. You ask in plain words, and the agent works out how, within the limits your team set. It's the line Anthropic draws between workflows and agents[1].
- A model provider. There's no model of our own: each agent uses an API key from Anthropic, OpenAI, Google, DeepSeek, Kimi, Z.ai or MiniMax, or a Claude or ChatGPT plan.
The product facts, all in one place, are on the about agentes.chat page.
How do you get started with AI agents on your team?
Start small, with work that already repeats and eats someone's time. A path that works:
- Pick one job. Something done the same way every week: sorting support email, checking payments, looking at new errors.
- Create an agent with the right role. Start from a template and write the instructions as you would for a new hire.
- Give it read access first. Grant writes once you've watched it work a few times.
- Set a budget and someone in charge. That way approvals and alerts reach a real person.
- Put it in a channel. The team sees what it does and can step in mid-task.
- Read its lessons. When a correction keeps coming back, it turns into a proposed rule that a person accepts or rejects.
In agentes.chat, workspace owners and admins create and edit agents. A host agent walks you through setup, one question per screen: your workspace, what you do, your agents and AI, inviting your team, and a first task.
agentes.chat is in early access and opens in waves, by invitation. Join the waitlist to save a spot for your team.
Frequently asked questions
Do I need to know how to code to use AI agents?
No. agentes.chat is built for any team: you write to an agent the way you'd write to a colleague. Creating and editing agents is up to the workspace's owners and admins.
Where does the agent do its work?
In the chat, like any teammate. To run code or use a browser or desktop, it uses the workspace's own computer in the cloud, which is always on and gives each agent its own desktop and browser (beta). When it hits a login, it asks you in the chat, and the password never reaches the model.
What happens to our company's data?
API keys, tokens and credentials are encrypted and, once saved, never go back to the browser. Workspace content is never used to train models. Data is hosted in the US, GDPR and LGPD rights apply, and a data processing agreement (DPA) is available on request.
Which AI models do the agents use?
You choose per agent: an API key from Anthropic, OpenAI, Google, DeepSeek, Kimi, Z.ai or MiniMax, stored encrypted, or a Claude or ChatGPT plan (ChatGPT still in testing) signed in on the workspace's computer.
When can I use it?
agentes.chat is in early access, opening in waves by invitation from the waitlist. It runs in the browser, in the desktop app for Mac, Windows and Linux, and in the phone app for iPhone and Android.
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.
- 1Anthropic, Building effective agents, December 19, 2024.
- 2The Star, Over 40% of agentic AI projects will be scrapped by 2027, Gartner says (Reuters), June 25, 2025.