Agentic AI

Agentic AI describes systems that can plan multi-step tasks and work on them using available tools.

An agent selects the next step, executes it and evaluates the intermediate result. This repeated action loop is the defining feature; the same language model may simply answer individual requests in a different application.

The typical structure of an agent

An agent typically combines four components: a language model for decisions, callable tools, stored task and history context, and a loop for working through the task step by step.

After each step, the system checks whether the goal has been reached, further work is needed or a stopping condition applies. Tools may be connected through tool use or MCP.

Why this matters for engineering teams

A coding agent can read a task, search the codebase, modify files, run tests and revise its work based on the test output.

Teams need to define which actions are permitted without confirmation, which systems the agent may access and how changes are reviewed before production use.

The risks that come with independent action

Independent actions require clearly bounded permissions. Otherwise, an agent may modify data, interact with external systems or take other steps that are difficult to reverse.

Suitable guardrails include permission boundaries, confirmation for critical actions and logging of tool calls. These controls should be established before production use.