A chatbot responds to a request. An agent is designed to pursue an objective across several steps. It may inspect a CRM record, search a knowledge base, draft a reply, update a ticket and ask a person for approval. The model supplies flexible reasoning; the surrounding system supplies permissions, state and constraints.
The agent loop
A robust loop begins with a goal and context, creates a plan, selects a tool, observes the result and decides whether to continue. Tool descriptions should be narrow and typed: “create refund up to £500” is safer than “manage orders.” The system should record every action, input and output so an operator can reconstruct what happened.
Memory is not one feature. Short-term memory holds the current task. Long-term memory may store preferences or facts, but it needs expiry, provenance and correction. Retrieval brings relevant documents into context; it should not be treated as proof that the documents are accurate.
Where agents work well
Good early use cases are bounded, repetitive and observable: triaging support requests, preparing sales research, reconciling routine documents, monitoring incidents and assembling internal reports. They are weaker where objectives are ambiguous, data is sparse or an incorrect action is expensive. A human approval step is appropriate for payments, contractual commitments, access changes and customer promises.
Controls before autonomy
- Use least-privilege credentials and separate read from write tools.
- Set budgets for time, tokens, API calls and transaction value.
- Require confirmation when an action is irreversible or externally visible.
- Test prompt injection through documents, emails and web pages.
- Measure completion rate, escalation rate, error severity and operator time.
Build or buy?
Managed agent products offer speed, hosted connectors and standard observability. Building internally offers more control over state, identity and domain-specific evaluation, but creates a maintenance obligation. A useful compromise is to buy commodity orchestration while keeping policy, data boundaries and approval logic in your own system.
Agentic systems should be introduced as operational software, not magical employees. Start with a narrow job, define its authority in writing, expose its uncertainty and expand only when evidence shows that reliability exceeds the cost of supervision.