Agent Orchestration in Maestro Processes
AI agents invoked inside UiPath Maestro processes: a process step starts an agent, the agent returns structured output, and that output is mapped to process variables used to route subsequent process execution such as RPA workflows, API workflows, connector activities or human approval tasks.
Recorded characteristics
- Function
- UiPath documents agents in Maestro as Service Tasks in BPMN processes: when the process reaches the Service Task, Maestro invokes the agent with input parameters for that process instance, the agent returns output data, and that output is used to determine the next steps in the workflow. Service Task implementation types include "Start and wait for agent" (deployed UiPath agent) and "Start and wait for external agent" (agent hosted outside UiPath, with CrewAI, Salesforce and ServiceNow given as examples). The same process model can also start and wait for RPA workflows and API workflows, execute connector activities, create queue items, evaluate business rules and create Action Center tasks. In the Maestro Flow canvas, an inline Autonomous Agent node is configured with a model, system prompt, user prompt and declared output variables; its structured output is referenced as $vars.<agent>.output.<field> and consumed by downstream Decision or Switch nodes to select a branch.
- Data access
- Agent inputs are the process-instance values passed into the agent through Service Task inputs or, in Flow, through prompt expressions referencing upstream process variables. An Autonomous Agent can additionally be connected to a Context resource pointing at a context index in the organisation, with configurable retrieval mode, query, result count, relevance threshold and citations. Built-in tools can analyse supplied files, transform a CSV row by row and summarise documents. Maestro documentation states inputs and outputs are passed by reference and governed by platform services such as Orchestrator, Integration Service and Data Fabric. Agents can be configured to report the tools used, decisions made and trace data for debugging and transparency. UiPath documentation reviewed does not establish any broader standing data access for the orchestration layer itself.
- Actions
- Can take actions
- External actions
- Conditional
- Human confirmation
- Conditional
- Permission basis
- Mixed
- Administrative control
- Documented controls include: designing the BPMN process or Flow and choosing each task or node implementation; selecting the agent, external agent, RPA workflow, API workflow, connector activity or business rule a task calls; mapping inputs, outputs and process variables; configuring gateways, decisions and branches; adding human approval through User tasks/Action Center or through an Escalation resource on an Autonomous Agent; agent-level model, system prompt, user prompt, temperature, maximum tokens, maximum iterations (1-100, default 25) and guardrails; publishing the process as a solution version and deploying it to Orchestrator, which requires the Packages - Create permission or an Automation Publisher role; configuring execution identity for child processes through folder isolation or explicit Package Requirements execution settings (runtime type, account including "Run as myself", and machine); upgrading a deployed process to a new published version; and instance management plus monitoring of running instances.
- Default state
- Disabled
- Availability
- UiPath lists agents, external agents, Integration Service agent connections, human-in-the-loop, action tasks, API workflows, instance management, debugging/simulating/upgrading/publishing and the monitoring dashboard as available on both Automation Cloud/Test Cloud and Automation Cloud Dedicated/Test Cloud Dedicated. The feature availability page notes that use of external agents is supported only for Google Vertex, Azure Foundry, Databricks and Snowflake, and that access to agents integration and API workflows depends on the cloud offering. Some adjacent items are marked preview, including Start and wait for Function, Execute business rule, Case Manager Agent and building a Maestro Case with a coding agent.
- Licensing
- UiPath states Maestro requires a combination of platform, user and consumption licences, with entitlements depending on the Unified Pricing or Flex plan, and that the same Maestro licensing applies to Maestro Case, Maestro Flow and Maestro BPMN processes. Work run inside a Maestro instance consumes the native consumables of the task types used. A debug run does not consume Agent Units or Platform Units for process execution, although services the debug run calls consume their own units.
- External model or provider
- Not established for the orchestration layer itself. Maestro is documented as vendor-neutral orchestration; in the Flow Autonomous Agent node the model is selected per agent from the models available to the organisation, defaulting to the organisation default agent model. No specific model or provider is named for Maestro orchestration in the documentation reviewed.
- Limitations and uncertainty
- Authority is mediated: the agent returns output, and any persistent operational change is performed by the configured downstream task (RPA workflow, API workflow, connector activity, queue item, message) or by a tool connected to an Autonomous Agent. UiPath documentation reviewed does not establish that the orchestration layer grants agents arbitrary system access. External agents retain their own authority and controls; invoking CrewAI, Salesforce or ServiceNow through Maestro does not transfer those systems' authority to UiPath. Human approval is optional and depends on process design, so unattended end-to-end execution is possible. Identity is documented at the child-job level (Orchestrator account and machine, folder identities, Integration Service connections, Action Center users) rather than as a single documented agent identity, so permission basis is recorded as mixed. Agent Builder authoring, Action Center itself, Autopilot and the Conversational Agent node are treated as separate matters and are not part of this record.
Evidence
- Maestro - Using agents in Maestro
Supports: Function · Data access · Limitations · Primary source
Agents are invoked as Service Tasks with instance input parameters and return output that determines next steps.
Agents receive per-instance input parameters and can report tools used, decisions made and trace data.
Documentation describes agent output driving next steps rather than the orchestration layer performing actions itself.
- Maestro - Service task
Supports: Function · Actions · External actions · Permission basis · Limitations · Primary source
Service task implementation types for deployed agents, external agents, RPA workflows, API workflows and connector activities.
Process steps can start RPA workflows, API workflows, connector activities, queue items and notifications following an agent step.
Start and wait for external agent triggers an agent hosted outside UiPath; connector activities run Integration Service operations against external systems.
Child agent or RPA process runs as a separate Orchestrator job using any available identity in the deployment folder unless execution settings specify otherwise.
External agents are hosted outside UiPath and retain their own execution environment.
- Maestro - Tasks
Supports: Function · Primary source
Task types map to platform components including agents, robots and Action Center.
- Maestro - Orchestrate an AI agent
Supports: Function · Actions · Default state · Primary source
Worked example of an agent decision routing downstream execution.
Agent output routes execution into downstream automated steps.
A flow must be built and triggered before an agent step executes.
- Maestro - Autonomous Agent node
Supports: Function · Data access · Actions · External actions · Human confirmation · Admin controls · External model · Primary source
Inline Autonomous Agent node with prompts and declared structured outputs.
Context resource grounds the agent in a context index with configurable retrieval; built-in tools analyse files, transform CSV rows and summarise documents.
Agents can call connector tools, RPA workflow and API function tools and MCP server tools, and decide during reasoning whether to call them.
Connector tools can, for example, send an email or create a record in a downstream system.
Escalations handle is optional and routes cases the agent should not decide alone to a person before the agent resumes.
Agent model, prompts, temperature, max tokens, max iterations and guardrails are configurable on the node.
Model is selected per agent from models available to the organisation, defaulting to the organisation default agent model.
- Maestro - Integrating systems and data
Supports: Data access · Permission basis · Availability · Primary source
Inputs and outputs are passed by reference and governed by Orchestrator, Integration Service and Data Fabric.
Runtime data flow is governed by Orchestrator, Integration Service and Data Fabric.
Access to agents integration and API workflows depends on the cloud offering.
- Maestro - Maestro feature availability
Supports: External actions · Availability · Limitations · Primary source
External agent support is limited to Google Vertex, Azure Foundry, Databricks and Snowflake.
Feature availability across Automation Cloud, Test Cloud and dedicated offerings, including preview items.
Several adjacent items, including business rule execution, Functions and Case Manager Agent, are marked preview.
- Maestro - User task
Supports: Human confirmation · Admin controls · Primary source
User task pauses process execution until a human assignee completes the task.
Human approval steps are configured in the process model with assignees and forms.
- Maestro - Human node
Supports: Human confirmation · Primary source
Human node outcomes such as Approve and Reject route subsequent execution.
- Maestro - Publishing, deploying, and upgrading agentic processes
Supports: Permission basis · Admin controls · Default state · Primary source
Child processes run as Orchestrator jobs under a folder identity by default, or under an explicitly configured account and machine, including Run as myself.
Publishing permissions, deployment to Orchestrator, execution settings and version upgrades.
A process must be designed, published as a solution version and deployed to Orchestrator before it can run.
- Maestro - Licensing
Supports: Licensing · Primary source
Platform, user and consumption licences; task types consume their native consumables; debug runs do not consume Agent or Platform Units for process execution.