Skip to main content

Core Concepts

UiPath's Agent Builder framework is built around five core components: generative AI models, prompts, tools, context, and escalations. Knowing what each one does makes the build walkthrough much easier to follow.

System prompt​

The instructions that define the agent's purpose, tone, and boundaries — similar in spirit to a job description. This is the first thing you write when designing an agent in Studio Web.

Tools​

Tools are the actions an agent is allowed to take during execution: existing UiPath automations, Integration Service activities (pre-built connectors to systems like Salesforce or SAP), or even other agents it can call. Adding a tool doesn't mean the agent will always use it — the agent decides, at run time, whether and when a tool is relevant to the case in front of it.

Context grounding​

Context grounding is how you connect an agent to your business data so its decisions are based on real information rather than the model's general knowledge alone. You link one or more Context Grounding indexes — permissioned knowledge bases — in the agent's Contexts section, and the agent retrieves relevant data from them as it reasons through a case. Good context grounding is often the difference between an agent that gives generic answers and one that gives answers grounded in your actual policies and data.

Escalations​

Escalations are UiPath's mechanism for human-in-the-loop control, powered by Action Apps in Action Center. You use escalations for two situations: asking a human to validate case details before the agent runs a sensitive tool, or asking for manual help when the agent hits a business exception it shouldn't decide alone. Designing good escalation points is as important as designing the agent's happy path.

Agent memory​

Agents can retain relevant information across a conversation or across steps within a single case, so they don't need everything re-stated at every turn — this is what allows multi-turn, conversational interactions rather than single-shot Q&A.

The build-test-publish loop​

Once system prompt, tools, context, and escalations are configured, you test the agent using the built-in Debug chat (which supports multi-turn conversations), then publish it as a solution package to Orchestrator, where it can be invoked on its own or wired into a larger Maestro process.

Next: Build your first agent puts these pieces together.