Agentic AI and the New Software Stack: What Changes When Models Can Act

01 FROM ANSWERS TO ACTION

A conventional language-model interaction ends with text. An agentic system adds a loop: it interprets a goal, selects a next step, uses an available tool, observes the result, and decides whether another step is warranted. The distinction matters because action introduces state, permissions, and the possibility of effects outside the conversation.

02 THE STACK EXPANDS

The new stack is not one magical model. It is a coordination layer around a model: a planner or policy, tool adapters, a state store, an execution runtime, and checks that can stop or redirect work. Anthropic's discussion of agent patterns is useful here because it separates predictable workflows from systems that let a model dynamically choose steps.

03 PLANNING IS A CONTROL PROBLEM

Planning can make a system more useful, but every extra step creates another opportunity for a wrong assumption. A robust design makes the goal and the allowed actions explicit, keeps tool results legible, and asks for confirmation before consequential operations. NIST's risk-management framing reinforces that capability and trustworthiness have to be considered together.

04 MEMORY CHANGES THE STATE

Short-lived context is different from durable memory. A transcript can help a model continue a task; a persistent record can influence later decisions and therefore needs provenance, retention rules, and a way to correct errors. Treating memory as a product surface—not an invisible prompt trick—makes ownership and deletion questions answerable.

05 EVALUATION FOLLOWS TRAJECTORIES

A final answer is not enough to evaluate an acting system. Teams also need to inspect the path: which tools were selected, what data was exposed, where the system retried, and whether it stopped when it should. Scenario tests, traces, and human review can reveal failures that a polished final response hides.

06 GOVERNANCE IS RUNTIME LOGIC

Permissions, audit logs, rate limits, sandboxing, and human approval are not paperwork added after deployment. They are runtime components of an agentic product. The practical shift is from asking whether a model is smart to asking whether the entire loop is bounded, observable, and recoverable when the model is wrong.