What is agentic software? How AI agents work and where they’re useful

You ask software to prepare a supplier comparison. Usually, that means opening tabs, downloading documents, copying prices into a spreadsheet and discovering that one supplier has hidden the delivery charge on page twelve.

Agentic software aims to take on more of that work. Given a goal and suitable access, it can gather information, decide what to investigate next, use other applications and work towards a result. You still define what you want and what it is allowed to do. The software takes responsibility for more of the steps in between.

That is the appeal of agentic AI: being able to delegate a piece of work without specifying every click.

What is agentic software?

Agentic software uses AI to choose and carry out steps towards a goal, drawing on information and tools and adjusting its approach as results come back. Its independence is bounded by the way the system is designed and the permissions it receives.

In MIT Sloan’s “Agentic AI, explained”, Beth Stackpole describes AI systems that can perceive, reason and act, operating with varying levels of human supervision. Connecting AI to other software lets it participate in a workflow, from collecting information to taking an authorised action.

The terminology is still developing. “AI agent” commonly describes an individual system performing a task, while “agentic software” can describe an application incorporating that behaviour. Some researchers use “agentic AI” more specifically for systems involving several agents. For this article, we use the broader meaning: software in which AI has some control over how work proceeds.

Autonomy does not imply consciousness, personal intentions or unrestricted access. It describes the decisions a system can make within its assigned task.

How is it different from a chatbot or ordinary automation?

The chat window can be misleading. Two applications might look identical while giving the AI very different responsibilities.

A basic chatbot could explain how to compare suppliers. An agent-enabled application could retrieve their quotations, notice that one omits a delivery date, search for the missing information and prepare a comparison with unresolved questions clearly marked.

Traditional automation can also execute actions. The architectural distinction is who determines the route. Anthropic’s engineering guide distinguishes workflows with predefined paths from agents that dynamically direct their processes and tool use.

For the supplier comparison, a basic chatbot could explain the criteria or analyse quotations you paste into the conversation. Fixed automation could retrieve specified fields from known systems and fill a template. Agentic software could choose which available sources to inspect, follow up on gaps and adapt the comparison to what it finds.

These approaches can coexist. A purchasing application might use an agent to investigate an unusual quotation, ordinary code to calculate totals and a fixed approval process before an order can be placed.

The agentic part is the discretion to choose the next step. A product label alone tells you little about how much discretion exists.

How does an agent work?

OpenAI’s practical guide describes three basic components: a model, tools and instructions. The model interprets the task and helps decide what to do. Tools provide access to information or actions. Instructions define the job and its constraints.

Those components operate in a loop: inspect the available context, choose a next step, use a tool, examine the result and continue or stop. A system may use one agent or coordinate several, but multiple agents are not a prerequisite.

For our supplier comparison, a tool might return a quotation with an unfamiliar unit of measurement. The next useful step would be to resolve that unit before comparing prices. If it remains ambiguous, the system could flag the uncertainty for a person.

That ability to respond to intermediate results is what makes the workflow adaptive. Whether it responds correctly is a separate question that testing must establish.

What could you use it for?

The potential applications extend across work that involves interpreting information and moving between systems. These examples illustrate possible designs; they are not claims about products we have deployed.

Research and analysis. An agent could investigate a market question, search for supporting sources, follow conflicting findings and assemble a report. The output should make it possible to inspect where its conclusions came from.

Software development. An agent could investigate a bug, inspect relevant files, propose a change and run tests. A failed test supplies information for the next attempt; code review provides a separate decision about whether to accept the change.

Evidence review. An agent could compare a set of requirements with supporting documents, locate missing records and prepare a list of unresolved conditions. A reviewer would still need to distinguish evidence that was not found from evidence that demonstrates a failure.

Business operations. An agent could help prepare a supplier assessment or assemble the documents needed to process a customer request. Permissions could allow investigation and drafting while reserving commitments or account changes for approval.

A useful design starts with a defined outcome. “Prepare a comparison covering price, delivery and warranty, with sources and unresolved questions” gives a system a clearer task than “handle procurement”.

Why businesses are interested

MIT Sloan identifies an economic opportunity in reducing the effort involved in searching, communicating and coordinating work. An agent may create value by improving a decision, or by making a sufficiently useful decision cheaper and easier to reach.

That matters because a great deal of work happens between applications. Someone reads a document, checks another system, asks for clarification and updates a record. Each action may be small, but completing the sequence requires attention.

Agentic software offers a way to delegate some of that coordination. Its value depends on the result and the cost of getting there, including the human effort required to review it.

Time saved also needs careful interpretation. As MIT Sloan’s article points out, freeing part of someone’s working day does not automatically produce the same percentage reduction in labour costs. The benefit might instead be faster service, greater capacity or more attention for difficult cases.

Where the promise meets the paperwork

Connecting a model to a business process creates implementation work. Data needs to be accessible and interpretable. Actions need permissions. Someone needs to own the system when a task stalls or a result is disputed.

MIT Sloan describes one healthcare implementation in which 80% of the work involved activities such as data engineering, governance and workflow integration. That figure belongs to the reported case; it is not a universal estimate. It does illustrate why selecting a model is only one part of deployment.

The risks grow with the actions an agent can take. An incorrect recommendation can mislead a reader; an incorrect account update changes the record itself. Reliability, access control and accountability therefore belong in the design from the beginning.

For an evidence-review assistant, we would start with permission to inspect documents and draft findings. We would measure missed requirements, unsupported conclusions and the time reviewers spend correcting the output. Broader responsibilities would depend on those results.

When is an agent worth using?

Anthropic recommends starting with the simplest approach that meets the need. A predictable task with clear rules may be better served by conventional automation. Agentic behaviour becomes more interesting when the next step depends on interpreting changing or incomplete information.

Our suggested starting point is a bounded task with an observable result and a manageable cost of error. Give the system representative cases, including awkward ones, and compare it with the existing process. Count the checking and repair work as well as the work it completes.

For the supplier comparison, the useful test is whether the resulting document lets someone make a decision with less effort and acceptable accuracy. Check the units, follow the citations and look for that delivery charge on page twelve.

Sources

• Beth Stackpole, “Agentic AI, explained,” MIT Sloan.
• Anthropic, “Building effective agents”.
• OpenAI, “A practical guide to building agents”.

The examples and proposed evaluation approaches are illustrative analysis by AgenticSoftware.

Cover photo: Karl Pawlowicz on Unsplash.