Agents
What is an AI agent, and how is it different from a chatbot?
An AI agent is software that uses a large language model to decide what to do next and then does it, by calling tools such as a CRM, an inbox or a database, until a goal is reached. A chatbot answers in words; an agent takes steps in real systems.
Updated · 5 min read
How does an AI agent work?
Almost every AI agent runs the same loop. The model reads the goal and everything it has learned so far, picks a tool and the arguments to call it with, and waits. The software around the model runs the tool and hands the result back. Then the model decides again: call another tool, or write the final answer.
- Read the request, the instructions and any earlier results.
- Choose a tool, such as “search contacts”, and fill in its arguments.
- The platform checks the call, runs it with stored credentials and records it.
- The result goes back to the model as new information.
- Repeat until the goal is met, or until a limit stops the run.
The model never touches your systems itself. It only asks for a tool call. Whether that call is allowed, which credentials it uses, whether a person must approve it first, and what is written to the log are all decided by ordinary software. That is why the platform an agent runs on matters as much as the model.
What is the difference between an AI agent and a chatbot?
A chatbot turns a message into a reply. An agent turns a request into a series of actions, and the reply is only the last step. The difference shows up in what can go wrong, and so in what you need around it.
| Chatbot | AI agent | |
|---|---|---|
| What it produces | A reply in words | Actions in other systems, then a reply |
| Tools | None, or a few lookups | Reads and writes through connected tools |
| Steps per request | One | Many, chosen as it goes |
| Worst mistake | A wrong answer | A wrong action, such as an update to the wrong record |
| Controls it needs | Good content | Permissions, approvals, limits and an audit trail |
What it produces
- Chatbot
- A reply in words
- AI agent
- Actions in other systems, then a reply
Tools
- Chatbot
- None, or a few lookups
- AI agent
- Reads and writes through connected tools
Steps per request
- Chatbot
- One
- AI agent
- Many, chosen as it goes
Worst mistake
- Chatbot
- A wrong answer
- AI agent
- A wrong action, such as an update to the wrong record
Controls it needs
- Chatbot
- Good content
- AI agent
- Permissions, approvals, limits and an audit trail
The line is not sharp. A chatbot that can look up an order status is already calling a tool. What matters is whether the software can change something outside the conversation. Once it can, you need controls a chatbot never needed.
What are the parts of an AI agent?
- Instructions: what the agent is for, how it should work, and what it must never do.
- A model: the language model that reads, plans and chooses tools. See how to choose a model for an agent.
- Tools: the actions it may take, each with a name, a description and the arguments it accepts.
- Limits: how many steps, how much time and how much money one run may use.
- Approvals: the actions that must wait for a person to say yes. See human-in-the-loop approvals.
- A record: every model call and tool call, so you can see what happened and why.
What are AI agents good at?
Agents do well when the input is messy, the next step depends on what they find, and there are clear tools to act with. Typical examples:
- Qualifying a new lead: look the company up, create or update the contact in the CRM, and post a short summary to the sales team.
- Answering visitors on a website, including questions that need a lookup, such as an order or a booking.
- Sorting a shared inbox: read each message, decide what it is, and draft a reply or hand it to the right team.
- Pulling a report together from several systems that have no shared view.
They do less well when a mistake cannot be undone and nobody reviews it, or when the same steps run the same way every time. For that, a workflow is cheaper and more predictable: see AI agents vs workflows.
How do you build an AI agent in MyDeepAgents?
In MyDeepAgents an agent is a definition you edit in the console: instructions of up to 200,000 characters, a model with optional fallback models, and up to 64 tools. Tools come from four places: actions of the connectors you have connected, tools of MCP servers you have added, other agents in your workspace, and agents at other companies reached over the A2A protocol.
Each run is durable. It runs as a Temporal workflow, so a run waiting for a person or a slow tool keeps its place even if a server restarts. Each run also has limits. By default a run may make 40 model calls, use 10 minutes of active time and spend $0.50 on the model. When it reaches a limit it stops with the status “stopped by limit”, and you can raise the limits per agent.
You try a draft in the console’s test chat. Every message there is a full run, and the chat shows each tool call with its input and output. Publishing creates a numbered version that never changes. The staging and live names point at versions, so going back to an earlier version means pointing live at it again.
An agent can also call another agent of the same workspace as a tool. The other agent runs as a separate run, with its own tools and limits, by default up to three levels deep.
Key points
- An AI agent uses a language model to choose and call tools until a goal is met.
- A chatbot produces words; an agent changes things in other systems.
- The software around the model decides what is allowed, runs tools with stored credentials and records every call.
- Give every agent limits on steps, time and cost, and approvals on the actions that matter.
- When the steps never change, use a workflow instead of an agent.
Questions people also ask
Is every chatbot built on a language model an AI agent?
No. A chatbot that only answers in words is not an agent. It starts to act like one when it can call tools that change things outside the conversation, such as updating a record or sending an email.
Does an AI agent learn from each conversation?
Not by itself. The model does not change between runs. In MyDeepAgents an agent does not keep a memory between runs yet: each run starts from its instructions, plus the earlier messages of the same conversation when there is one.
What stops an AI agent from running forever?
Limits set outside the model. In MyDeepAgents every run has a cap on model calls, active time and model cost (40 calls, 10 minutes and $0.50 by default), and a run that reaches one ends as stopped by limit.
Can an AI agent use another AI agent?
Yes. In MyDeepAgents an agent can call another agent of the same workspace as a tool, and that agent runs with its own tools and limits. Agents at other companies can be called over the open A2A protocol.