Agents in Flow (Preview)

UiPath Agents in Flow lets you embed low-code agents directly inside a Maestro Flow project as inline agent nodes. Instead of building an agent separately in Agent Builder and invoking it through a process call, you configure the agent's system prompt, model, input/output schema, and tools from the same Studio Web canvas where you design the flow.

Inline agent nodes are first-class nodes on the flow canvas. They appear alongside connectors, script nodes, and other flow components. Flow variables pass to and from inline agents natively, and you can chain multiple agents in series within a single flow.

Inline agents vs. standalone agents

	Inline agents in Flow	Standalone agents
Configuration space	Flow canvas in Studio Web	Agent Builder in Studio Web
デプロイ	No separate publish step — the agent is part of the flow	Must be published as a standalone agent process before use in a flow
Variable exchange	Flow variables referenced natively in the agent prompt and output	Variables passed through process invocation parameters
Chaining	Multiple agents in series in a single flow	Invoke one agent from another using the Agent tool
Variable insertion in prompts	Uses $ conventions ($vars, $metadata)	Uses @ conventions from Studio Web

Agent types

• Autonomous — The agent selects and executes tools to complete a task.

ツール

Inline agents support the following tool categories:

ツールのカテゴリ	利用可能状況
RPA processes	利用可能
API ワークフロー	利用可能
Activities (アクティビティ)	利用可能
Other flows (Maestro)	利用可能
Agent (invoke a published standalone agent)	利用可能
Integration Service connectors	利用可能
Built-in tools (Analyze Files, Summarize, Batch Transform)	利用可能
エスカレーション	利用可能
MCP tools	利用できません。
A2A (Agent-to-Agent)	利用できません。
IXP (document extraction)	利用できません。

Context

Context grounding is available for inline agents at parity with Agent Builder.

ガードレール

You can add guardrails at the agent level, LLM level, or individual tool level. Guardrails evaluate agent inputs and outputs against defined policies and trigger an action on a match.

Built-in guardrails detect specific content categories automatically:

• PII detection
• 有害コンテンツ
• Intellectual Property detection
• User prompt attacks

エスカレーション

Escalations are available for inline agents at parity with Agent Builder.

Output variables

Every inline agent returns an output object. The default field is content, which holds the agent's response text. Additional output variables can be declared in the Variables tab to return structured data — each variable requires a name and type, and the system prompt should direct the agent to populate them.

Output from upstream nodes is referenced in downstream nodes using $vars.<node>.output.<field>, for example $vars.agent1.output.content.

Debugging and evaluations

デバッグ

You can debug individual inline agent nodes in isolation from Studio Web using breakpoints, step-by-step execution, and execution traces. Individual agent nodes can be debugged in isolation by selecting the node and using the Debug step-by-step option.

Full flow runs appear as jobs in Orchestrator, where you can monitor execution status. After a run completes, the execution trace is available in the Execution tab of the bottom panel, showing a step-by-step breakdown with timing for each node including LLM calls and outputs.

評価

You can evaluate inline agents per-node from the canvas using the Evaluate button in the toolbar. The Datasets and Evaluators tabs in the bottom panel manage evaluation datasets and evaluators and add evaluators to a dataset run. The Runs tab shows evaluation run history.

Evaluation for inline agents is broadly at parity with Agent Builder, although the experience for creating evaluations differs. Some Agent Builder capabilities are not yet available for inline Flow agents. For details, see Limitations.

Three evaluator categories are available:

カテゴリ (Category)	入力	説明
Output Validation	Contains Text	Passes when the agent output contains the specified text
Output Validation	Exact Match	Checks whether the agent output exactly matches the expected output
Output Validation	JSON Similarity	Compares the structural similarity of JSON objects or values
LLM Judge	LLM Judge Output	Uses a language model to score how closely the output matches the expected output in meaning and intent
LLM Judge	LLM Judge Strict JSON	Uses a language model to judge strict JSON structural match
LLM Judge	LLM Judge Trajectory	Uses a language model to judge whether the agent followed expected behavior
LLM Judge	LLM Judge Simulation	Uses a language model to judge the simulated agent execution path
Tool Call	Tool Call Arguments	Checks that tools were called with correct arguments
Tool Call	Tool Call Count	Checks the number of times each tool was called
Tool Call	Tool Call Order	Checks that tools were called in the expected sequence
Tool Call	Tool Call Output	Checks that tool outputs match expected values