Loading
An AI agent takes a goal, chooses its own steps, calls tools and checks the result.
Hassan Sarwar ·
An AI agent is a program that is given a goal rather than a script. It chooses its own next step, calls tools such as an API or a database, observes the result and adjusts. That loop is what separates an agent from a model that only answers when prompted.
Everything else in the term is marketing. If a system cannot take an action in another system and cannot see what that action did, it is not an agent, however it is described.
Every working agent has the same four parts, and each one is a place where a project fails. The goal has to be narrow enough to check. The tools have to be real API calls with real permissions. The decision step has to be observable. And the feedback loop has to have a stopping rule, or the agent will keep going.
A plain model call is reactive. You send text, it returns text, and nothing in your business changed. An agent is the same model wrapped in permissions and a loop, so a decision turns into a write against your CRM, a ticket transition, or a message to a person.
That difference is entirely about blast radius. The moment a model can change a record, you need guardrails, an audit log and an escalation path, which is why agent projects are integration projects wearing a different hat.
The interesting engineering in an agent is not the prompt. It is the set of constraints: which records it may read, which fields it may write, what it must never do without a human, and what happens when it is unsure. Those rules are what make an agent safe enough to leave running.
In practice we write them down before the first tool is connected, because they decide the architecture. An agent that may only draft is a very different system from one that may send.
If the process is deterministic, write the automation. A rule engine that reorders stock below a threshold is cheaper, faster and easier to debug than a model deciding the same thing, and it never has a bad day. Agents earn their cost where the input is unstructured and the next step genuinely depends on reading it.
The other disqualifier is a missing record of truth. An agent that has to guess which of three systems is correct will produce confident, plausible, wrong actions. Fix the data question first.
Start with one workflow, in draft-only mode, with a human approving every action for the first few weeks. Log everything. When the approval rate stops moving, you have an evaluation set and an argument for widening the permissions. That progression is what our Integration Ladder describes: assisted, then automated, then agentic.
No discovery marathon. Bring one process that costs your team hours and we will map what an integration would look like in a free 30-minute working session.