What is an AI Agent and some quick tips for better results

I was watching a video about identity and access management (YouTube link but be warned it's a bit technical) and I came across one of the best definitions of what an AI Agent is.

Quoting from the talk:

More than a chatbot, an autonomous software entity that perceives, decides, and acts.

It's simple and it's quite to the point.

KEY CHARACTERISTICS

Autonomous Execution: Operates independently with varying degrees of human oversight

Tool Usage: Connects to databases, APIs, MCP servers, and external services

Memory: Learn from past interactions, maintain context across sessions, and execute multi-step tasks reliably.

Multi-Step Reasoning: Breaks complex tasks into sub-tasks, sometimes delegating to other agents

Amazing overview.

My advice

Please consider that this is my advice and you may have another opinion and experience. That's fine.

Don't make it 100% autonomous

It should be as constrained as possible so that it doesn't do anything dangerous or break things. Also it should follow your rules and should do things your way. Blindly accepting the agent's results is a recipe for failure.

Connect to tools with care

There are a lot of horror stories online of companies where the agents deleted a database and you don't want that. Making things safe is important.

I understand that sometimes it's better to connect to everything that you have and tell Claude "check everything and...", but doing this is also dangerous. If Claude goes down a strange path, it could do something that you don't want it to.

It's the same thing as inviting someone to your organization and giving them permission to do everything. Super dangerous, not to mention the privacy implications.

Agents are amnesic

They forget everything each time you start, so saving memories is important. I don't like the way Claude, for example, keeps memories of everything automatically, so I like to control what to keep and what to discard. Not keeping a memory is also useful in certain cases where you can wipe the session and start fresh, especially when the agent goes in a crappy direction.

Multi-step reasoning

I always start with this: state the problem, define the solution before starting, adjust where I think it's incorrect, and proceed with my plan. I would advise you to do the same. Be in control always.

There are many more things that the agent can do like having skills, tools, etc., but I think keeping things simple is always the best approach. Skills are useful in very defined contexts so use them with care.

No feature is a silver bullet, at least in my opinion.

Final Thoughts

I will write more about this in the future, but I wanted to share this quickly because I think it's quite useful.

Photo by Immo Wegmann on Unsplash