Agents
What is an AI agent?
An AI agent is a program given a goal rather than an instruction. It decides the steps itself, calls tools to carry them out, reads the results and iterates until the goal is met. A chatbot answers; an agent acts.
Updated 9 August 2026
The questions people ask
What is the difference between an AI agent and a chatbot?
A chatbot produces text and stops there. An agent holds tools and uses them: it opens a document, updates a record, sends a reply. The difference is not the model's intelligence, it is the right to act you grant it.
What is an autonomous AI agent?
An agent that chains its steps without approval at each one. Autonomy is a dial, not a state: you choose what it does alone and what still needs a human. In practice reading delegates easily, sending outward much less so.
Can an AI agent get things wrong?
Yes, which is why the question is not how to stop it but how to make its mistakes repairable. Every action traced, every write reversible, and the sensitive moves put to a human before they run.
What tools does an AI agent use?
The ones you declare to it. A search across your data, a document read, a record write. The MCP protocol standardised that declaration, so an agent can now use a third party's tools with no bespoke development.
We probed 485 domains to find which ones actually expose such tools: 142 answer, 55 of them without OAuth.
Do you need to code to build an AI agent?
Not for the part that decides the outcome. What makes an agent good is the precision of its goal, the exact list of its tools and the limits you set. That is written in plain language, not in Python.
How to build an AI agent
Four decisions, in this order. Taking them backwards is the main reason agents run without producing anything useful.
- 01
Write the goal, not the procedure
«Match every incoming invoice to the right customer record» beats ten detailed steps. The agent will find the procedure; the goal is the part it cannot invent.
- 02
Pick its tools, and only those
An agent with access to everything wanders and becomes impossible to audit. Give it document reading and record writing, nothing more if nothing more is needed.
- 03
Set the limit before it goes live
Decide up front what ships without approval and what stops at a human. Adding the limit afterwards always means adding it too late.
- 04
Watch it work on real cases
An agent is judged on your data, not on a test set. The awkward cases are the ones that reveal whether the goal was well written.
What an agent is never left to do alone
The part nobody writes about, and the only one that matters once the agent runs on real data.
- Writing outward: an email that leaves, a published document, a payment. Those stop at a human.
- Deleting. A deletion cannot be replayed, so it cannot be delegated.
- Stepping past the permissions of the person it works for. An agent inherits its principal's rights, never more.
- Acting without a trace. Every tool call goes to the audit log, with its author and the chain that triggered it.
What it looks like in production
At Tiepolo, agents ingested 440,000 documents in two days: reading the scans, extracting the data, creating linked records. At Lilycare, they open insurer portals and produce quotes. In both cases a human approves what goes out.
Agents in Standards
They work inside your business objects, under your teams' roles, with the audit log and human approval switched on by default.