Start here
Agentic AI in plain language
The important distinction is control of the next step. A predefined workflow can contain several model calls without being meaningfully agentic. In an agentic design, observations influence a model-led choice about what to do next.
Autonomy is a design choice, not an all-or-nothing feature. A system can investigate independently yet require approval before changing a customer record. More autonomy is useful only if the added flexibility outweighs extra cost, delay and failure modes.
How it works
Separate predictable work from uncertain work
Keep straightforward validation and routing in ordinary code. Identify the part where the next useful action genuinely depends on what is found.
Define decisions and boundaries
List permitted actions, evidence requirements and stopping conditions. Identify actions that must wait for a person.
Compare against the simpler baseline
Use the same test cases for the agentic system and a fixed workflow. Compare task completion, incorrect actions, elapsed time and execution cost.
Illustrative example · Fictional data
Handling a delivery exception
- Input
- A fictional customer says a parcel has not arrived. The system can read tracking events, check the delivery address and draft a support response.
- Process
- If tracking shows a failed attempt, request clarification. If tracking is unavailable, record the failure and escalate. If delivered, ask for confirmation rather than automatically issuing a refund.
- Expected output
- A supported next-step recommendation, with the relevant tracking event and an unsent draft response.
- How to check it
- Test contradictory tracking events and missing addresses. The system should not treat a plausible explanation as proof or turn investigation permission into refund authority.
Where it helps
- Exception handling with several possible investigation paths.
- Research planning when each finding changes the next question.
- Developer workflows that adapt to test feedback rather than following a fixed repair script.
Choose a workflow before choosing a framework
- Fixed workflow
- Use known steps for predictable inputs. It is easier to reproduce and inspect.
- Agentic workflow
- Let the system choose some next steps when the route cannot be usefully predetermined.
- Multi-agent design
- Divide work among agents only when distinct responsibilities justify coordination and handoff costs.
Planner, tools, state and memory
A bounded agent combines a decision-making component with an execution environment. The planner proposes the next step; tools perform permitted reads or actions; state records progress, observations and pending approvals. Memory can retain selected information across steps or runs, but must have a defined purpose, retention policy and access boundary.
For a return request, an agent might read the order, retrieve the applicable policy, compare eligibility and draft a response. An approval gate can stop it before issuing a refund or sending the email. A failed lookup should lead to a retry within a limit or an escalation—not an invented order status. Set maximum steps, time and spend as well as the goal.
Agentic workflows, RPA and framework choice
Rule-based automation suits stable, explicit procedures. An agent can help interpret variable text or choose among permitted steps, but introduces uncertainty that deterministic automation may avoid. A practical system can combine both: use a model to classify a request and conventional code to validate and execute the transaction.
A framework is optional. Compare state persistence, retries, tracing, human approval and deployment requirements before choosing plain application code or an agent framework. Retrieval is useful when a task needs external knowledge; neither RAG nor one reasoning pattern is compulsory for every agent.
Related reading and learning pathways
Common pitfalls
- Calling a workflow agentic does not explain which decisions are delegated; document them.
- Additional agents can repeat errors or lose context during handoffs.
- A model evaluating its own answer is not independent proof that the task succeeded.
Put the idea to work
Try it yourself
Map a real recurring task into fixed steps and uncertain branches. Design an agent only for one branch, then describe a test where a fixed rule would outperform it.
Common questions
Is agentic AI the same as generative AI?
No. Generative AI creates content. Agentic AI concerns selecting actions towards a goal. An application may use both. For example, a model that drafts an email is generative, while a system that decides to search a calendar, book a slot and then send that email is acting agentically.
Does agentic AI require a particular framework?
No. The control pattern can be implemented with application code or a framework. Choose based on state, testing, integration and operational needs. Many teams prototype the loop in plain code first. A framework becomes worthwhile when you need persistent state, retries, human approvals or tracing across many steps.
Is a multi-step prompt automatically agentic?
No. Following a predefined list of steps differs from choosing actions based on external results. Describe the actual behaviour rather than relying on the label. A useful test is to ask whether the next step can change based on what a tool or environment returned. If it cannot, the flow is a fixed pipeline.
Your next step
Continue learning
Connected concepts
Sources and further reading
Examples on this page are teaching exercises using fictional information. They are not client case studies or measured outcomes.