- Getting started
- Prerequisites
- Building agents
Agents user guide
Escalations and Agent Memory
As a developer, you may want to give your agent a way to validate business case details with a human before running a tool, or to request manual assistance if they run into issues. In other cases, there may be business exceptions in a process when you want an agent to have a human make an approval or some other type of decision.
This is where escalations come in. Escalations are powered by Action apps in Action Center and give developers the tools to design and configure human-in-the-loop events in agent execution.
Agent Memory is a service inside each agent that helps it remember facts and observations as it works. We support Agent Memory through:
- Active memory: Consists of an aggregated log of all of the processes and tasks that an agent executes during its runtime. These are automatically retained by agents as chunks of action-observation pairs in Context Grounding indexes.
- Long-term memory: Consists of a combination of available knowledge bases (through Context Grounding), reflections on active memory observations, and the implicit agent memory itself.
Active memory is implicit in agent design and helps track escalations to reduce repeated human intervention on the same task types.
You can enable Agent Memory in the Escalations panel. Once enabled, Agent Memory is tied to the latest version of the published agent. At runtime, this version stores by default store each Action Center task triggered by an agent escalation. Users resolving the escalation also have the option to not store the escalation in agent memory.
You can set up long-term memory using Context Grounding indexes. Context Grounding allows you to create an index of your stored documents that the agent can query to find supporting or relevant information which may help it during its run. For details, refer to About Context Grounding.
You can create multiple escalations for different situations where an agent may require human assistance.
To help the agent differentiate these escalations, configure the following fields for each escalation:
- Description: This helps agents identify the escalation and when to use it.
- Assign to: The Action Center assignee. You can assign an escalation to a user in your UiPath® organization.
- App templates: Select the Action app you want to use to surface the escalation. You must first deploy an app to your tenant before you can access it from the Agents workspace.
- Outcome behavior: Specifies how each outcome should be handled by the agent after an escalation was resolved.
- App input fields: If the selected app has input properties marked as required, these will be displayed here. Describe the content of the field, so that the Agent can infer its value and fill it in when raising the escalation.
- Agent Memory Enabled: Activate this option to store question and response pairs from escalations, and use text embeddings to automatically resolve escalations with content that is very similar to past escalations. Agent Memory stores design time and runtime escalations separately.
After you create an escalation, you can reference it from the system prompt. Use explicit instructions, such as: "Before using Tool X, be sure to raise the Human Confirmation escalation and include the details for the action you're proposing".
Agents do not currently support the In/Out argument type for Action Schema properties.
