What makes an AI system agentic?

Short answer

On this site, an agentic AI system means a model embedded in a multi-step process with instructions, state, tools and an environment. Its actions can change the state presented at the next step, so the relevant unit of analysis is the interaction trajectory rather than an isolated response.

Model

Instructions

Context

Memory

Tools

Environment

Action

Changed state

new context

Everything above the action shapes it; the action changes the state; the changed state comes back as new context.

For a simplified one-shot analysis, a model call can be treated as a mapping from supplied input and configuration to a distribution of possible outputs. An agentic application adds orchestration around that call. Instructions tell it what role to play. Context carries the conversation and whatever documents or tool results have been loaded into view. Memory persists things across sessions. Tools let it act: run code, edit files, place orders, send messages. The environment is whatever those actions touch.

The feature relevant to this research is the multi-step feedback loop. Each action can alter the environment or the record of the interaction, and those changes can influence later actions. The system’s behaviour is the whole trajectory, not any single reply.

A concrete example

A coding assistant with repository access works through an afternoon. By the twentieth step, its context holds its own earlier edits, failing test output and tool results. Its next action depends on all of that accumulated state. The same model, asked the same question in a fresh session, might do something quite different.

A common misunderstanding

“The model is the system.” The model is one component of the measured system. Two copies of the same model with different instructions, tools, memory, orchestration or history are, for measurement purposes, different systems. Evaluating the model alone leaves most of the machinery, and most of the behaviour, unexamined.

Why it matters

Safety and reliability properties can live in the loop rather than in the model: in what accumulates, what persists and what feeds back. That is why Overdog measures whole interactions under controlled conditions rather than scoring single responses.