Exécuter des agents

Unified runtime is the execution foundation that runs all agents, regardless of how they are built. It provides a single runtime model that ensures agents behave consistently whether they are created using low-code tools or written directly in code.

At a high level, unified runtime provides:

• One execution engine for all agents
• One runtime contract that defines how an agent runs
• One operational model for debugging, monitoring, and resuming execution

Agents are represented using a standardized definition that includes prompts, tools, inputs, outputs, and settings. At runtime, this definition is interpreted and executed by the unified runtime engine.

This enables a "best of both worlds" approach:

• Low-code users can build and iterate quickly using visual tools
• Pro developers can extend or convert agents into code when needed
• Enterprises benefit from consistent governance, security, and observability

This eliminates fragmentation between low-code and coded agents and avoids the need to rebuild agents when moving between development styles.

Unified runtime execution model

When an agent runs:

1. The agent definition is prepared as a self-contained execution contract.
2. The runtime executes the agent in a secure, serverless environment.
3. Long-running steps (such as human-in-the-loop actions or external processes) can pause execution.
4. The agent resumes automatically when the required action completes.
5. Execution details are captured through traces and logs.

All of this happens using the same runtime, regardless of whether the run is triggered from Debug or from Orchestrator. From a user interface perspective, unified runtime is largely invisible.

From Agent Builder and Orchestrator:

• You select Debug to test an agent.
• You publish the agent to a UiPath tenant as a solution and run it as an Orchestrator job — triggered via API, a Maestro node, or a process activity.
• You view inputs, outputs, tool calls, and traces.
• You see agent runs move through familiar states: running, pending, suspended, completed, or failed.

Traçages

Because all agents run on the same runtime, traces have a consistent structure regardless of whether the agent was built with low-code tools or written in code. The traces UI surfaces:

• Appels et réponses LLM
• Tool invocations and results
• Errors and guardrail actions

Related topics

• Running an agent
• Gérer les agents UiPath
• Traçages d’agent