What agentic means

The word agentic is on most vendor slides, applied to everything from a renamed chat product to systems that work through a task for hours without anyone watching. The meaning worth keeping is narrower, and it is older than the current products. Russell and Norvig’s standard AI textbook has defined an agent since its first edition in 1995 as anything that perceives its environment and acts on it.¹ The AI models behind today’s chat products are large language models. By that definition an agent is a large language model with hands. It can use tools, it reads the result of each step and keeps going until the task is done, and it knows when to stop and ask.

How an agent differs from a chat assistant

In a chat, the AI model reads a message and writes a reply, and the person does everything around it. The person opens the intercompany agreement, copies the relevant clause and last year’s margins into the chat, asks for a view, pastes the answer into the memo, notices that the margin it quotes does not match the segmented accounts, and asks again.

An agent runs the same model, given tools and a goal. Asked to prepare the response to a tax authority’s question about why a distributor’s margin fell below the range set in the group’s policy, the agent opens the agreement itself, reads the policy, pulls the segmented profit and loss account from the reporting data the team has given it, computes the shortfall against the range, checks whether a year-end adjustment was booked, drafts the factual part of the response with the evidence attached, and checks the draft against the numbers before reporting back. It stops where a judgment call is due, for example which business reasons to put forward and how much of the explanation to volunteer, and asks. The hands, the loop of acting and checking, and the checkpoints where it stops and asks make the difference.

A Harvard Business Review field study put the distinction plainly.² The assistant handed people information to act on, while the agent handed them completed work. Anthropic’s engineers use a shorter definition.³ An agent is a large language model autonomously using tools in a loop. It is the same definition, stated as a mechanism. The loop is the part that is easy to miss. The model acts, reads the result, decides the next step, and goes again until the task is done or it reaches a checkpoint.

What an agent is made of

Opened up, an agent has four parts and a loop. The parts are the model, the instructions, the tools, and the memory. The loop ties them together and stops at checkpoints where a person decides.

The AI model is the engine. It reads what is in front of it and decides the next step. The model keeps nothing from one session to the next, which is why memory is a part of its own.

The instructions tell the agent what the task is for and how the organization does the work. They also say what is off limits. In a transfer pricing team, that ranges from the purpose of the work to the house view on method selection and the cases that always go to a person.

The tools are the hands, from reading and writing files to searching, running a calculation, or opening an application such as the mailbox or the finance system. Each tool is a capability the organization has decided to give the agent. For anyone weighing risk, the model never executes anything itself. It asks for a tool, the platform it runs on executes the action and returns the result, and the team decides which tools the agent may use at all.

The memory is what the agent has in front of it for the task at hand and whatever it keeps across sessions. How much survives from one session to the next is a matter of setup, the subject of our article on long-horizon agents.

In the loop the agent gathers context, acts, checks the result, and goes again until the task is done or it reaches a checkpoint, where a person decides. The checkpoints are part of the design, set by the team that runs the agent, and the person can step in at any point.

In practice the model and most of the hands come from the agentic AI platform the organization already runs, such as Claude Cowork or Copilot Cowork. The platform supplies the model, the general tools for files, the web, and applications, and the memory. On top sits a specialized layer made of the instructions, the domain tools, and the checkpoints that turn a general agent into a transfer pricing agent. A team can write that layer itself or take it from a specialist. It is the layer we build at Supernomial.

Degrees of agentic

Agentic comes in degrees. Andrej Karpathy, one of the field’s best-known researchers, calls the range an autonomy slider.⁴ At one end the agent suggests each step and a person accepts or rejects it. At the other it takes the whole task and reports back. Most real work sits in between and moves along the slider as the team’s trust grows. A transfer pricing team starts with an agent that drafts and checks under supervision and widens the agent’s autonomy, task type by task type, as results prove reliable. Some definitions reserve agentic AI for systems of several coordinated agents.⁵ In everyday use the word describes a single agent that acts, which is the sense used here.

The definition doubles as a test when a vendor uses the word. Three questions settle it. Can the system use tools on the team’s behalf, does it keep going on its own until the task is done rather than waiting for a new prompt at each step, and does it know when to stop and ask? A product that answers no to all three is a chat window with a new name. One that passes the first two and fails the third is an agent the team cannot yet trust. Gartner’s advice sorts work along the same slider, with assistants for retrieval, automation for routine tasks, and agents where decisions are needed.⁶

For a transfer pricing team the question has moved from the model to the setup. McKinsey expects the scale of agent adoption to be capped by how much oversight people can provide.⁷ What remains to decide is which hands the agent gets, which instructions it works from, and at which points a person signs off.

References

1. Russell, S. and Norvig, P. (1995) Artificial Intelligence: A Modern Approach. Englewood Cliffs, NJ: Prentice Hall. 4th edn (2021), Hoboken, NJ: Pearson. Chapter 2. Available at: aima.cs.berkeley.edu. ↩
2. Yang, J., Zyskowski, K., Yonack, N. and Ma, J. (2026) ‘Research: How AI Agents Broaden the Scope of Knowledge Work’, Harvard Business Review, 29 July. Available at: hbr.org. ↩
3. Anthropic (2025) ‘Effective context engineering for AI agents’, Anthropic Engineering, 29 September. Available at: anthropic.com. ↩
4. Karpathy, A. (2025) ‘Software Is Changing (Again)’, keynote at Y Combinator AI Startup School, San Francisco, 17 June. Available at: youtube.com. ↩
5. OECD (2026) The agentic AI landscape and its conceptual foundations. OECD Artificial Intelligence Papers No. 56. Paris: OECD Publishing, February. Available at: oecd.org. ↩
6. Gartner (2025) ‘Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027’, press release, 25 June. Available at: gartner.com. ↩
7. McKinsey & Company (2025) ‘The agentic organization: Contours of the next paradigm for the AI era’, 26 September. Available at: mckinsey.com. ↩