- Getting started
- Prerequisites
- Building agents
Agents user guide
About agents
Agents represent the next generation of automation. They are well suited for non-deterministic, dynamic use cases, and leverage large language models (LLMs) to plan, think, take action, and learn over time.
Agents are goal-based, act independently, and can make dynamic decisions. They are best suited for ad hoc tasks that require high adaptability. Agents are creative and intuitive, they have decision making skills, and can handle ambiguity. Most importantly, agents have AI skills which enable them to communicate in natural language, plan out the steps to accomplish a task, and coordinate with other robots and agents for process-level outcomes. Refer to the Capabilities section for more details on what agents are good at and when they are not the right fit.
Robots, on the other hand, are rule-based, act predictably, and are suitable for deterministic actions. Robots are best used for routine tasks that require high reliability and efficiency. They are structured and logical, and work on efficiency-oriented systematic processing.
Agents and robots can work together to solve end-to-end business problems and to enable enterprise-grade agentic automation. Agents handle tasks that robots may not be able to, whereas robots help institute control, determinism, and governance for agents as they operate.
An agent consists of four core components:
- Natural language prompt: Instructions or
plan for the agent, that determine its role, goal, and constraints.
- Prompts, made of input and output arguments, are of two types: user prompts and system prompts.
- Context: Knowledge bases that the agent can use to find information. Context is based on:
- Short and long-term memory.
- Context Grounding, which represents the ability to query from knowledge-bases to ground prompts.
- Agent memory, or how the agent learns based on its runtime.
- Tools: What the agent uses to take action. The agent invokes tools based on the prompt.
- Available tools: activities, automations, micro automations, or other agents.
- Escalations: The human in the loop.
- Agents can involve a human when necessary, to help gather additional information or review arguments.
- Agent escalation paths: Action Center action apps, communication channels.
- Communicating: Natural language collaboration between agents and users.
- Initiating: Triggered by user requests, but also by system events.
- Planning: Understands and plans out the tasks to complete a process.
- Deciding: Can independently make dynamic decisions.
- Adapting: Access to enterprise context, apps, and systems.
- Healing: Knows when a workflow is broken and works to resolve it.
- Learning: Remembers preferences, previous requests, and actions.
- Coordinating: Works with robots, other agents, and users.
Not all tasks or operations are well-suited to agent use cases. Agents built on LLMs can help orchestrate simpler workflows. Relying on them for complex data processing or multi-step logic without additional specialized infrastructure can lead to frequent failures.
Good use cases for Agents: Smaller, well-defined tasks with limited context and lower risk, such as summarizing short texts, drafting emails, etc. Each of these narrowly scoped agents can then be orchestrated together for more effective and reliable output.
Bad fits for Agents: Large or complex data processing, high accuracy requirements, and tasks involving high stakes (financial, legal, or regulatory) where consistency and zero error tolerance are paramount.
