Large language models are increasingly capable of handling complex, multi-step tasks, giving rise to a new category of LLM-powered systems known as agents. This guide, drawn from numerous customer deployments, is designed for product and engineering teams building their first agents. It covers frameworks for identifying promising use cases, patterns for agent logic and orchestration, and best practices for keeping agents safe, predictable, and effective.
What is an agent? While conventional software helps users streamline workflows, agents perform those workflows on the user's behalf with a high degree of independence. A workflow is a sequence of steps needed to reach a goal, such as resolving a customer service issue, booking a reservation, committing a code change, or generating a report. Simple chatbots, single-turn LLMs, and sentiment classifiers that do not use the LLM to control workflow execution are not agents.
An agent uses an LLM to manage workflow execution and make decisions. It recognizes when a workflow is complete, corrects its actions when needed, and can halt and hand control back to the user if it fails. It also has access to tools that let it gather context and take actions in external systems, selecting the right tool based on the workflow's current state and always operating within clearly defined guardrails.
When should you build an agent? Agents suit workflows where traditional rule-based approaches fall short. Payment fraud analysis illustrates the difference: a rules engine works like a checklist, while an LLM agent evaluates context and subtle patterns, identifying suspicious activity even when no clear rule is broken. Prioritize workflows involving complex decision-making, rules that are difficult to maintain, or heavy reliance on unstructured data. If your use case does not clearly meet these criteria, a deterministic solution may suffice.
Source: OpenAI Blog · Summarized by HeadlinesBriefing