Skip to content

AI Agent

An AI agent is a system that uses an AI model to plan or decide how to complete multi-step work, often with access to tools or external systems.

An AI agent is a system that uses an AI model to plan or decide how to complete work across one or more steps, often with access to tools or external systems.

The defining feature is not simply that it can chat.

It can participate in a workflow and choose actions based on the task.

OpenAI's current agent documentation describes agents as systems that can plan and complete tasks using tools, maintain context across steps, and coordinate work. OpenAI agent documentation

What an Agent Can Include

An agent may combine:

The model provides flexible decision-making inside that system.

Agent vs. Chatbot

A chatbot may simply:

Receive message → generate response.

An agent might:

Receive request
   ↓
Determine required steps
   ↓
Search records
   ↓
Call API
   ↓
Evaluate result
   ↓
Request approval
   ↓
Perform action
   ↓
Report outcome

That ability to choose and sequence actions introduces additional capability and risk.

Agent vs. Workflow Automation

A Workflow Automation follows a predefined process.

An AI Agent can choose among paths dynamically.

For example:

Workflow

If invoice status = overdue → send reminder.

Agent

Review account history → decide which information is relevant → choose an appropriate supported action.

The deterministic workflow is usually preferable when the rule is already known.

Permissions Matter

An agent with Tool Use may be able to:

  • Send email
  • Change CRM records
  • Issue refunds
  • Publish content
  • Search private data
  • Modify files

Those permissions should follow the same least-privilege thinking used in Authorization and Administrative Access.

Why It Matters

Agentic systems change AI from:

Producing information

toward:

Taking action.

That makes failures potentially more consequential.

Common Problems

Common problems include:

  • Calling a chatbot an agent
  • Agent has administrator-level access
  • No approval for irreversible actions
  • Prompt Injection
  • Poorly defined stopping conditions
  • Agent loops
  • No logging
  • Difficult-to-reproduce behaviour
  • Tool failures interpreted incorrectly

How to Improve

Start with limited authority.

Define:

  • Allowed tools
  • Permissions
  • Maximum scope
  • Human approval thresholds
  • Logging
  • Failure behaviour
  • AI Evaluation

Expand autonomy only when testing justifies it.