In this article
We all know what a software application is. You download and install software on your computer, and with the help of your operating system (Windows, iOS or Linux), that software opens and runs. Some applications receive their instructions from a computer – for example loading your mouse and keyboard when your system starts. Other applications must be started and managed with human input – for example connecting a computer to a new WiFi router.
The software provides a framework to manipulate digital information which could be words, numbers, images, CAD drawings, music, video or more. A human decides what the software will do next by typing in text from a keyboard, clicking buttons with a mouse or swiping with their fingers. The software has capabilities, it has rules, it has security, and it can even use other digital tools: send a document to a printer, or play music on a headset. Complex custom software applications can drive a car, or instruct a vacuum cleaner, or fly a rocket.
And now there are AI Agents. We are learning about AI as a tool for writing better emails, but where does an AI Agent fit in with what we already know?
Think of a traditional software application and an AI agent as two different kinds of employees.
A traditional software application
A traditional app is like a junior grad. They’re trained for specific tasks and are superbly good at those tasks – fast and consistent. But if the task moves beyond their their knowledge, skills and experience they freeze or make mistakes. It’s up to you as the boss, to retrain them for new scenarios. Or hire a new employee with a different skillset.
An AI Agent
An AI agent is like an experienced employee who can think on their feet. You don’t give them step-by-step instructions for every scenario, they figure out the steps themselves, and ask a question if they aren’t sure.
- You provide them with context – goals and access to tools (like the internet, or coding software).
- They can read what you give them, and also research similar options you didn’t provide.
- They decide what approach to take based on what they learned.
- They can use tools and often the tools they are given are traditional software applications we already know. For example an AI agent can look something up in a database, check a supplier list, send information to another system or flag something for review.
- They adapt: If they receive an invoice from a new vendor they know to look for an invoice number, date and amount, using general knowledge of what invoices look like. Every new style of invoice they see, add to their knowledge.
- They can self-correct: “That looks like a date but is it month first or day first — let me check with the boss.”
How else are AI Agents similar to humans
- If you’ve been working on a task all week, you will remember some information without having to look it up. But if you only do this task once a month, you might not remember the right next step, or details like a bank account number. If you have to remember too much, you need to write notes. AI also has short term memory, but needs assistance with long term memory. That’s why it gives priority to instructions at the beginning and end of long prompts, and skips something important you added in the middle.
- Both humans and AI agents can chain together multiple steps and decisions on its own to work towards a goal. Which means it can overshoot, get distracted, and go down rabbit holes.
- Like humans, AI agents can make mistakes around what they predict, and they can overstep their authority in trying to reach the goal.
Mitigating the human-like risks of AI Agents
Memory limitations
Give the AI agent external reference material rather than relying on it to remember. Allow it to reference the same written spec every time. External databases, lookup tables, and written procedure notes remove the need to rely on recall. They also reduce the use of expensive tokens by keeping the context windows small.
Overshooting, distraction, and rabbit holes
Constrain the AI agent’s scope and tools. If you only give an agent three tools — read a PDF, extract fields, flag for review — it physically cannot go down a rabbit hole into your accounting system or start drafting emails.
You also limit the number of steps or iterations it’s allowed to take before it must return a result, the way you’d tell a human “spend no more than five minutes on this and come back.” Breaking large goals into smaller, well-defined sub-tasks also keeps the agent focused — “extract these six fields from this invoice” is hard to derail; “process this invoice end to end” is easy to derail.
Unpredictability and mistakes
The “human-in-the-loop” review is critical. The agent doesn’t need to be perfect — it needs to be good enough that a human reviewer can confirm or correct quickly. The human must assume the AI will be wrong sometimes and make catching those moments fast and easy. You also mitigate mistakes by restricting what the agent can do with its output automatically.
Overstepping authority
The agent should be able to read documents and suggest extractions, but not delete files, modify configurations, or push data to an ERP without approval. The principle is simple: give the agent enough autonomy to be useful, but not enough autonomy to cause damage.