Genesys

Tools in AI Studio wrap data actions that retrieve live information or perform transactions in connected services. Builders select an action, define required inputs and link values from start context or other tools. Precondition checks run before execution, with builder rules or Python expressions for validation. Output mappings and error codes control how the agent responds after a successful call, authentication failure or service error.

Sourcehelp.genesys.cloud/408957read 2026-10-20

A large action model selects tools and plans several steps toward a customer goal. Architect can combine that agent with conventional flow logic through the Call Agentic Virtual Agent action. The calling flow receives end context and separate success, failure and timeout paths. Exit reasons distinguish errors, escalation, customer requests and guardrail events, allowing the flow to route the next stage of the interaction.

Sourcehelp.genesys.cloud/409099read 2026-10-20

The Knowledge tab connects an agent to a knowledge base or knowledge configuration and shows its connected sources and synchronization information. Builders describe the material's scope, set retrieval conditions and link required context variables. Input validation can restrict when a search runs. Separate outcome instructions handle an unanswered query and the follow up after a match, while the retrieval language follows the selected resource's capabilities.

Sourcehelp.genesys.cloud/409012read 2026-10-20

Built in protection rules block prompt injection, role manipulation and harmful requests before input reaches the agent. Administrators add custom boundaries, violation limits and exit responses. An escalation tool transfers the conversation to a person, inheriting whether escalation is enabled from the calling flow. These controls interrupt or redirect an interaction; the agent has no staff approval inbox for reviewing and releasing individual proposed actions.

Sourcehelp.genesys.cloud/408935read 2026-10-20

Agent, version, job and session operations each carry their own AI Studio permissions, with separate add, edit and view rights, so authoring, publishing and session review can sit with different roles. Previewing an agent additionally needs integration viewing and voice bot connection permissions. Genesys Cloud holds a SOC 2 Type 2 attestation. Sensitive tool outputs and starting values can be masked so the action model never reads their raw contents while the data action still receives them.

Sourcehelp.genesys.cloud/articles/ai-studio-permissionsread 2026-10-20

Architect's execution history and replay expose the sequence of data action and knowledge calls made during an agent conversation. Each call carries latency and a status such as success, timeout, error or guardrail blocking, with inputs and outputs available for inspection. The same view records whether the interaction was contained, handed off or escalated and retains an exit reason, including a tool timeout or customer transfer request.

Sourcehelp.genesys.cloud/articles/test-and-troubleshoot-agentic-virtual-agentsread 2026-10-20

Start context passes participant data from Architect into the current agent interaction. Tool inputs and outputs can reuse those values, and end context returns selected information to the calling flow, up to 50 variables with string outputs capped at 2,000 characters. Once the interaction closes, anything worth keeping has to be written back to a business system through a data action, because the agent configuration has no store of its own that carries a customer's details into the next conversation.

Sourcehelp.genesys.cloud/409057read 2026-10-20

Customers select the Genesys Cloud region for their tenant and stored data. Core locations include Virginia, Oregon, Canada, Frankfurt, Ireland, London, Tokyo and Sydney, all on Genesys hosted AWS infrastructure. Satellite regions extend media handling closer to callers and can optionally store recordings, but cannot host an independent tenant. The European Sovereign Region is in limited availability. Every option is a Genesys operated cloud region, and no customer installed or on premises runtime exists for the agent.

Sourcehelp.genesys.cloud/articles/aws-regions-for-genesys-cloud-deploymentread 2026-10-20

AI Studio starts an agent with a customer supplied name, role and guidelines. The customer then assembles its tools, knowledge connection, start and end context, and guardrails before publication. A configuration can be saved while incomplete, but all required sections must be complete to publish it. Existing agents can be duplicated for testing; the creation path does not provide a selectable collection of finished service agents.

Sourcehelp.genesys.cloud/409064read 2026-10-20

Incoming voice and digital interactions can reach the agent through Architect bot flows with Virtual Agent enabled. A Call Agentic Virtual Agent action invokes the selected published agent automatically at the configured point in the flow. Each invocation uses the latest published version and passes literal or expression based start context. Success, failure and timeout routes continue processing after the agent returns.

Sourcehelp.genesys.cloud/409099read 2026-10-20

The Agentic Virtual Agent runs on a large action model that Genesys manages. When Genesys introduces a new model version, administrators get a transition window to duplicate an agent, test it on the new version in draft and choose when to promote it, after which the platform moves every agent to the latest version. The window governs timing only. Customers cannot pick a different foundation model maker, set a model per agent or bring their own model endpoint into the agent configuration.

Sourcehelp.genesys.cloud/articles/transition-to-the-latest-agentic-virtual-agent-modelread 2026-10-20

Public agentic virtual agent APIs expose the platform's own agent and tool contracts through the Developer Center. Tool types include data actions, knowledge bases and knowledge settings, with shared properties and subtype specific inputs. The API Explorer currently omits some inherited fields because of its OpenAPI 2.0 renderer. A separate specification supplies complete contracts while the reference moves to OpenAPI 3.0.

Sourcehelp.genesys.cloud/articles/agentic-virtual-agents-public-apisread 2026-10-20

Testing is manual. The AI Studio preview widget lets an author try the agent's role, guidelines, data actions, knowledge configuration, start and exit behavior and guardrails in a chosen language and voice or digital channel, then restart with the latest configuration.

Architect's digital bot simulator opens a test conversation with the author playing the customer; only published flows run there, so each change is republished before retesting, and external resource actions can behave differently there than in production. Replay mode supports troubleshooting after the fact. No saved test cases, scripted scenarios or scored runs exist for the agent.

Sourcehelp.genesys.cloud/308320read 2026-10-20

Service work runs through data action tools and knowledge lookups. Each tool has structured inputs, an output object and error handling, and Architect continues the workflow on the success, failure or timeout path after the agent returns. A data action can submit a transaction to any service with an API, but no tool opens a browser session, reads a screen or drives mouse and keyboard input, so a system reachable only through its user interface stays out of the agent's reach.

Sourcehelp.genesys.cloud/408957read 2026-10-20