Agentic AI is taking the cloud by storm. With cloud providers releasing their own agentic frameworks. AWS Has Bedrock AgentCore (which I experimented with during the 2025 Hackathon), GCP has one called Vertex AI, and Azure has Microsoft Agentic Framework. But what is Agentic work? When should you use it or not use it?
Let's start with what is Agentic. All Agents share the same high level features.
- Perception: They use data to identify problems. Like receiving input from an alert then looking up system specs in a database.
- Autonomy: Agents work without human input toward a goal. This could be completing a transaction, fixing a bug, or writing new code.
- Adaptability: Agents can respond to unexpected input or requests without needing to be told what to do.
Some people put in other parts like learning or collaboration. But I see these as sub sets to the main three.
We know what Agents are but where do we use them. Agents, as described above, are designed to be complex autonomous problem solvers. Which means they could, theoretically, be used everywhere and I have heard people say this. I strongly disagree.
If you have a simple problem, that is procedural, well defined, and repeatable you probably want to use a traditional programming approach. These problems are easy to code but can be harder to get an agent to format everything you want just so. These could make great tools for your agent but I argue they should not themselves be directed by an agent. Good examples of these are::
- Writing known formats to a database.
- Simple logic that doesn’t need to change.
- Verification of facts
- High stakes decisions
Agents thrive in the inverse of this. Especially when inputs are unknown and the logic is complex, but you have a clear goal or goals that you want to achieve.
- Customer service agents
- Programming - in particular drafting new features/prototypes
- Triaging
- When an action is better than no action
Agentic AI is a powerful tool in the developers belt now, but it is not a magic hammer that can fix every problem.