Optimizely Opal glossary

2 min

This glossary defines key terms and concepts used in Optimizely Opal. Use it as a reference when exploring Opal Chat, building agent workflows, or configuring Opal across Optimizely.

Action card – A structured Opal Chat response that supports in-line actions. Action cards include tables, graphs, forms, and images you can act on without leaving the conversation. See Action cards.

Agent – A specialized, pre-configured capability within Opal with clear skills and tools to handle specific requests or tasks. They can perform advanced actions, and you can include them in agent workflows. Examples include the following:

Campaign brief generation agent – Create high-quality briefs that are on-format.

Content translation agent – Convert existing content to a different language, considering cultural, company, and grammatical differences.

Agent card – An entry in the Agent Library that summarizes an agent with its title, emoji, description, tags, and ID.

Agent Directory – A central hub where you can browse, view, and install default agents for Optimizely Opal. See Agent directory.

Agent ID – The unique identifier for an agent, prefixed with @ to invoke it in Opal Chat, such as @blog_generation.

Agent Library – A browsable catalog of Opal agents on the Opal Chat home page, organized by use case. See Agent Library.

Agent workflow – A predefined, structured sequence of steps and decisions that a specialized agent within Opal follows to address a specific type of request or achieve a particular goal. Not the same as a workflow agent.

Anomaly rescue – Execution Advisor scenario that injects corrective guidance into runs that Execution Guardrails flag. Helps the run reach a valid result instead of terminating. See Anomaly rescue.

Artifact – Any canvas or file Opal creates in a chat thread, agent execution, or workflow run. See Artifacts page overview.

Ask Opal – A feature available in the global navigation bar of supported Optimizely products that launches Opal Chat for in-context support, guidance, or content generation. See Access Optimizely Opal Chat.

Baseline evaluation score – Lowest Evaluation Score that counts as Passed in Output Evaluation. Available values are 75%, 80%, 85%, 90%, and 95%. See Set the baseline evaluation score.

Canvas – An interactive document where you and Opal build, customize, and edit content together within a chat thread. See Canvas overview.

Connection – A link between Opal and a specific Optimizely product instance. A connection is required for Opal features to function within that product. You must create a connection before you customize Opal.

Connector tool – A system tool that connects Opal to a third-party service, such as Google Analytics, WordPress, or Microsoft 365. See Connector tools overview.

Context details – Any additional information provided along with your message that helps Opal understand the context of your request. This might include your active Optimizely product, the current page or URL, your organization ID, or any previously stored data relevant to the conversation. This additional context helps provide more personalized and relevant responses.

Conversation sample – Reference conversation that shows ideal multi-turn agent behavior. Output Evaluation uses conversation samples to score multi-turn agent runs. See Add a conversation sample for a multi-turn agent.

Credits (also called Opal Credits) – Units that track usage of Optimizely Opal's AI-powered features. You consume credits each time Opal makes an API call to an AI model, such as a large language model (LLM). This includes generating responses, running agent workflows, retrieving data, or performing AI-driven actions. See Optimizely Opal credits.

Custom tool – A tool built and maintained by your organization. See Custom tools.

Default agent – A ready-to-use agent built by Optimizely that helps you quickly get started with common tasks and workflows. See Default agents overview.

Enforcing mode – Execution Guardrails mode active after Learning and Watching modes complete. Opal stops runs that look clearly unsafe compared to learned behavior. See Enforcing mode.

Evaluation (Eval) – A structured assessment that measures agent outputs against quality criteria for accuracy, brand consistency, and relevance. See Evaluations in Opal overview.

Evaluation Score – Score Output Evaluation produces for a specialized agent run, between 0% and 100%. Opal compares the score to the baseline evaluation score to determine pass or fail. See How Output Evaluation scores runs.

Execution Guardrails – Sub-tab of the Quality tab that detects when a specialized agent run deviates from your agent's normal behavior. Stops unsafe runs automatically when configured. See Configure Execution Guardrails.

Execution Advisor – Runtime intervention layer for specialized agents. Recovers anomaly-flagged runs, retries failed tool calls, and reviews sensitive tool calls. See Execution Advisor.

Failure recovery – Execution Advisor scenario that generates recovery guidance when a tool call fails. Execution Advisor attempts to recover the run with adjusted input. See Failure recovery.

Global navigation bar – The horizontal menu available across Optimizely One products, which includes shortcuts such as Ask Opal and drop-down lists to change your organization and Optimizely product. See Overview of Opti IDOverview of Opti ID.

Guardrail Status – Outcome category Execution Guardrails assigns to a specialized agent run. Possible values are Passed, Failed, Warning, Terminated, Rescued, Unapproved, and Learning. See Execution outcomes.

History – Your past interactions within the current Opal conversation. For facts Opal stores about you across conversations, see Memory overview.

Instance skill – A skill that administrators set to apply organization-wide standards across an Optimizely product instance. See Skills overview.

Learning mode – Initial Execution Guardrails mode covering the first 20 successful runs of an agent version. Opal observes runs to build a baseline without blocking them. See Learning mode.

LLM-as-Judge – Large Language Model (LLM) that evaluates other model outputs against quality standards. Output Evaluation uses an LLM-as-Judge model to score specialized agent runs.

Memory – Facts Opal stores about you, including preferences, context, and working style. Opal retrieves memory automatically and applies it across conversations. See Memory overview.

Multi-turn conversation – An interaction mode in which a specialized agent stays active in Opal Chat after its first response, so you can give follow-up prompts and iterate on the output. See Specialized agents fundamentals and best practices.

New chat – A button in Opal Chat that clears the current session and starts a new interaction. See Opal Chat user interface. Examples include the following:

You want to discuss a completely new topic.

The conversation has gone off-topic.

You feel Opal is getting confused or stuck.

You want to ensure a clean slate for a critical query.

Notification Center – The full-page view where you browse, filter, and manage all notifications in Opal. See Notification Center.

Notifications tray – A compact view in Opal where you check recent notifications and take action without opening the Notification Center. See Notifications tray.

Opal App – The central interface for accessing Opal's features, including chat, skills, agents, tools, and connections.

Opal Chat – The conversational interface where you interact with Opal using natural language to ask questions, get insights, and perform tasks. See Optimizely Opal Chat.

Opal credits dashboard – The UI where administrators can monitor credit usage, allocation, and remaining balance. See View Opal usage.

Opal SDK – A way to simplify the creation of tools compatible with the Opal tools management service. See Create custom tools. Three SDKs are available, delivered through the following services:

Python through pip.

JavaScript (TypeScript) through npm.

C# through NuGet.

Optimizely Opal – Optimizely's agent orchestration platform that answers questions about the Optimizely platform, generates insights, and helps you complete tasks across Optimizely One. See Optimizely Opal overview.

Output Evaluation – Sub-tab of the Quality tab that scores specialized agent runs. Uses criteria, examples, and a baseline evaluation score you define. See Configure Output Evaluation.

Persistent chat – An Opal feature that preserves chat context as you move between supported Optimizely products.

Personal skill – A skill you create for your own preferences and working style. Personal skills apply only to your account. See Skills overview.

Preferred output – High-quality output you manually select from a specialized agent execution. Output Evaluation uses preferred outputs as benchmarks to score future runs. See Preferred output examples.

Product instance – A particular deployment of an Optimizely product within your organization's environment where Opal Chat is configured (for example, a specific Experimentation project).

Prompt – The natural language input you provide to Opal to initiate a request for information, assistance, or an action.

Quality tab – Section of the specialized agent builder containing automated quality scoring and runtime safeguards. Includes Output Evaluation and Execution Guardrails sub-tabs. See Quality tab overview.

Regenerate – The request to receive a different response to the same prompt. Use this option when you want alternative wording, phrasing, or ideas without changing your original query. See Opal Chat user interface.

Response – The output generated by Opal in reply to your prompt, which can include textual answers, suggested actions, data retrieved from Optimizely products, or interactive elements. See Action card for interactive responses you can act on in-line.

Response format – The specific way Opal presents the information it provides, dictating the structure and styling of Opal's output. Opal can present information in various formats, including the following:

Plain text – Simple, unformatted text.

Markdown – Advanced formatting, which can include rich text such as bold, italics, lists, headings, and tables.

HTML – Complex or specific web-friendly formatting.

Tables – Organized data in rows and columns.

Code snippets – Examples of code or scripts.

Retrieval-Augmented Generation (RAG) – A method where Opal retrieves your data and content before generating a response, so answers stay accurate and grounded. See Retrieval-Augmented Generation (RAG) overview.

Retrieval engine – The underlying system or component responsible for finding and fetching relevant information from its knowledge base and other data sources. See Retrieval-Augmented Generation (RAG) for the broader approach. Examples include the following:

Analytics retrieval engine – Retrieves quantitative data from analytics sources within Optimizely, such as experiment results, website traffic data, and conversion rates. Works in conjunction with the other components to provide data-driven insights and recommendations.

Knowledgebase retrieval engine – Searches Optimizely's documentation, tutorials, best practices, and knowledge base to return helpful articles, guides, and reference materials.

Opti data retrieval engine – Accesses your organization's configuration data within Optimizely, for example, experiments, feature flags, campaigns, or other Optimizely-specific configurations.

Sample skill – Ready-made skill examples you can copy and customize to quickly create your own skills. See Sample skillsSample skills.

Single-shot – The default interaction mode for specialized agents. The agent takes your inputs, runs once, and returns a final result. Use single-shot for automated workflows and tasks that do not require follow-up. See Specialized agents fundamentals and best practices.

Sensitivity – Execution Guardrails setting that controls how aggressively Opal flags or stops runs that fall outside the established baseline. The Sensitivity slider ranges from 1 (Lenient) to 5 (Aggressive). See Sensitivity.

Sensitive tool – Tool that performs destructive or high-impact actions, such as sending email, publishing content, or deleting data. Execution Advisor reviews sensitive tool calls before they run. See Sensitive tool approval.

Skills – Foundational context, rules, and behavioral guidelines that shape how Opal creates output and tailors it to your organization or personal preferences. See Skills overview.

Slack – A messaging platform that enables teams to communicate, collaborate, and integrate with tools like Optimizely Opal for real-time conversations and task management. See Slack.com.

Slack integration – A connection between Opal and Slack that lets you chat with Opal and complete tasks directly in Slack channels or DMs. See Optimizely Opal in Slack.

Specialized agent – A customized AI agent designed for specific, well-defined tasks. It uses targeted tools and defined inputs to finish work accurately. By default, these agents run once to give a result. You can also let them support multi-turn conversations to refine the output through chat. See Specialized agents overview.

System tool – A tool built and maintained by Optimizely. See System tools. See Connector tool for system tools that connect Opal to a third-party service.

Thread (also called chat session) – A continuous interaction with Opal Chat, consisting of one or more messages exchanged between you and Opal. Click New Chat to create a session. See Opal Chat user interface.

Thread ID – A unique identifier for a specific Opal chat session. To get your thread ID, see Copy your Opal Chat thread ID.

Tool (also called Opal Tools) – A discrete, API-exposed function or service that Opal can invoke to perform actions, retrieve information, or interact directly with Optimizely products and external services. System tools are built and maintained by Optimizely, and Custom tools are built and maintained by your organization. See Tools overview. Examples include the following:

Search web results and browse live URLs.

Take a screenshot of a webpage for analysis.

Upload files (video, image, .txt, .pdf, and so on) for context and analysis.

Tool manifest – A structured declaration (often a JSON object) that describes your tool to Opal. Think of it like a blueprint or a resume for your tool. It tells Opal what your tool does and how to use it. See Tool manifest.

Tool orchestration – The process where Opal selects, sequences, and executes its available tools in response to your request, ensuring the most effective and efficient completion of a task or retrieval of information.

Tone – A configurable option in Opal Chat that controls the style of the generated content, such as professional, casual, or technical. Useful when generating emails, documentation, or other written content. Change the tone Opal uses through skills.

Watching mode – Execution Guardrails mode that follows Learning mode and covers the next 30 successful runs. Opal flags unusual runs as warnings without blocking them. See Watching mode.

Workflow agent – An agent that automates a multi-step business process by orchestrating triggers, logic, and specialized agents. Not the same as an agent workflow. See Workflow agents overview.

Note

If you use Opti ID, administrators can turn off generative AI in the Opti ID Admin Center. See Turn generative AI off across Optimizely applicationsTurn generative AI off across Optimizely applications.