Optimizely

Integrations run in both directions across several classes. Inward, agents work natively across the Optimizely suite, CMS, Content Marketing Platform, Experimentation and Campaign, with system tools published per product, and Optimizely states Opal is powered by hundreds of tools.

Outward, pre-built integrations and MCP connectors reach Salesforce, Conductor and Google Analytics among others, and Optimizely's framing is that its agents sit on top of the customer's existing stack connecting third-party systems and Optimizely products alike, with no rewiring or tool swapping. Customers connect their own data and layer in custom instructions and tools, and administrators control which tools and registries are enabled. That spans CRM, analytics, SEO, content management and marketing execution.

Sourceoptimizely.com/products/ai and support.optimizely.com Opal documentationread 2026-09-02

Workflow agents are a documented construction. Optimizely's developer walkthrough sets out the three building blocks: a trigger that decides when the workflow starts, a series of agents that each perform one step, and logic determining how those steps are sequenced, with that logic expressed through conditions that branch the path and loops that repeat it. Steps pass structured output between one another, and the release notes describe Opal isolating required parameters and constructing the JSON handed from agent to agent.

A drag-and-drop builder assembles this without code, agents connect to existing data and tools across CMS, CMP, Experimentation and Campaign, and workflows carry an Active status and run unattended on their triggers. Vendor-reported platform data from May 2026 states nearly 1,700 companies have built over 4,000 agents run more than 172,000 times, with 32 percent of runs involving agents completing multi-step tasks through to completion rather than accelerating a single step.

Sourcesupport.optimizely.com workflow agent documentation, world.optimizely.com developer walkthrough and optimizely.com agentic momentum releaseread 2026-09-02

Grounding is a documented pipeline over the customer's own material rather than ad hoc retrieval. Before every model interaction the platform runs a multi-step enrichment pipeline that loads contextual information, and instructions are layered deliberately: instance-level instructions carry organization-wide brand voice and guidelines, personal instructions carry per-user context, and instance skills set organization-wide standards across a product instance.

Optimizely's framing on the product page is that great output starts with the right context and that agents are given skills, brand guidelines and compliance rules to work from. Agents also ground on the customer's own CMS and DAM content across the suite, and files can be attached directly to an agent, with the release notes recording support for CSV, PowerPoint and Word uploads.

Sourcesupport.optimizely.com Opal glossary and release notes, world.optimizely.com architecture walkthroughread 2026-09-02

Governance is layered, and a shipped mechanism enforces it. Execution Guardrails evaluate specialized agent runs and can pass, warn, terminate or mark them unapproved, with Learning mode covering the first twenty successful runs of an agent version so the system builds a baseline before it starts blocking, and Anomaly rescue letting an Execution Advisor inject corrective guidance into a flagged run so it reaches a valid result rather than dying.

Around that sit approval gates Optimizely states plainly, that teams review, edit and approve outputs before anything goes live and every workflow has room for approval; instance skills setting organization-wide standards; role-based permissions controlling who may modify those instructions; a tenant-wide switch disabling all Opal and generative AI features from Admin Center; new tools and registries disabled by default so an administrator must explicitly enable them; and compliance-checking agents reviewing copy for legal issues before publication.

Sourcesupport.optimizely.com Opal glossary and execution guardrails documentation, optimizely.com/products/airead 2026-09-02

Certifications, governance and controls are all documented. The certifications listed are ISO 27001:2022, ISO 27017:2015, ISO 27018:2019, SOC 2 Type 2, PCI DSS v4.0.1 and TISAX, with reports available on request, plus EU-US Data Privacy Framework certification and a documented HIPAA-ready CMS variant with a BAA.

A Trust Team with a Compliance Manager and a Director of Security Engineering reports to a CISO and is overseen by a Security Governance Board of senior executives, running periodic risk assessments on an ISO 27005:2018 methodology.

Authentication is required on all application entry points including the API, role-based permissions run across the platform, employee access is least-privilege through feature-limited portals with MFA and logged, and optional geofencing restricts which Optimizely staff may service an application by region.

Sourceoptimizely.com/trust-center compliance and security pagesread 2026-10-01

Per step execution logs record the reasoning, not just the result. Workflow agents keep an execution log showing the detail of every step, and a user can click into any step to see its input, its output, its processing status and its execution memory, which reconstructs a past run rather than reporting on a finished one.

The release notes add parameter generation logs to those execution logs: where agents make tool calls, the logs display within thinking chunks and capture how Opal interprets the input, isolates the required parameters and constructs the JSON passed between agents.

Alongside it, Execution Guardrails assign every specialized agent run a readable outcome, artifacts created in any chat thread, agent execution or workflow run are collected on an artifacts page, and employee access to customer applications is logged.

Sourcesupport.optimizely.com Opal release notes and workflow execution log documentationread 2026-09-02

A memory layer is documented with its own overview page. The glossary distinguishes two things by name: History, being past interactions within the current Opal conversation, and Memory, being facts Opal stores about the user across conversations, with a dedicated Memory overview in the documentation. The product page states the same in buyer terms, that memory builds over time so the platform gets smarter the more the team uses it. Separately, each workflow step carries its own execution memory visible in the run log, so state persists within a run as well as across sessions.

Sourcesupport.optimizely.com Opal glossary and Memory overview, optimizely.com/products/airead 2026-09-02

Optimizely lets the customer select the region. Its Trust Center publishes EU Data Hosting as an option a customer chooses, stating that where applicable the customer's data will be stored in EU data centers and that this gives full control over data location, framed explicitly for organizations with internal privacy and residency requirements.

Alongside it Optimizely offers Geofencing, a free service that applies regional controls to the Optimizely organization so that only EU-based support staff can service the customer's application, which addresses access as well as storage. The Services Description adds that hosting regions span Microsoft Azure, Google Cloud Platform and Amazon Web Services, that not every data center those providers offer is made available, and that the available data center for a given service is provided on request. EU Data Hosting is offered where applicable, so it is not available across every product in the suite, and which products qualify is not published.

Sourceoptimizely.com/trust-center EU Data Privacy and Optimizely Services Descriptionread 2026-09-02

A browsable Agent Library of more than fifty ready-to-use agents is documented as a product surface. The glossary defines an agent card as an entry in the Agent Library summarizing an agent with its title, emoji, description, tags and ID, which is a catalog with metadata a buyer browses and filters.

The published set covers task-specific roles including content translation, compliance and legal review, copywriting, personalization, content drafting and scheduling, competitive research and web performance analysis, and Optimizely states customers can use the out-of-the-box agents or build their own, with a no-code builder for the rest.

Beneath the agents sit skills, including instance skills that apply organization-wide standards across a product instance, and agent workflows as predefined structured sequences a specialized agent follows.

Sourcesupport.optimizely.com Opal glossary and optimizely.com/products/airead 2026-09-02

Optimizely documents four trigger classes in a dedicated help center article: a chat trigger running a workflow on demand from Opal Chat; a scheduler trigger starting it at a specific time or on a recurring schedule, with times set in UTC; a webhook trigger where Opal listens for an incoming call and starts the workflow when an external system sends an event, with configurable authentication headers and selection of the product instance to act against; and an email trigger firing when mail arrives at a nominated address, with sender and subject conditions to prevent spam starting a run.

Each trigger binds to a specific connected Optimizely instance. Channel and surface coverage is correspondingly wide: agents run natively across CMS, CMP, Experimentation and Campaign, are reachable from Opal Chat in the global navigation of supported products, and are callable from any MCP client.

Sourcesupport.optimizely.com workflow agent triggers and Opal release notesread 2026-09-02

No customer-facing model choice is documented. Opal is described as powered by large language models without naming a provider, and no model selector, per-agent model assignment, routing policy, bring-your-own-model or bring-your-own-key is published.

The controls that do exist are on or off rather than which: administrators can disable all Opal generative AI features tenant-wide from Admin Center, and new tools and registries are disabled by default until explicitly enabled.

An LLM-as-Judge model scores agent output, so more than one model role exists inside the platform, but the customer chooses neither; the Experimentation MCP server lets a customer's own assistant read Optimizely data.

Sourcesupport.optimizely.com Opal glossary and release notesread 2026-09-02

Optimizely offers several distinct routes in and out. It ships an Experimentation MCP server that authenticates through the customer's Opal instance and exposes linked experimentation data to any client supporting remote MCP, alongside a developer documentation site at docs.developers.optimizely.com, the Opal Tools SDK for building custom tools in Python and FastAPI, authenticated web and API entry points across the platform, and remote MCP connectors for reaching outward.

Customers build their own agents and register their own tools and registries, which administrators must explicitly enable, so third-party capability is a supported extension path with an admin control. The release notes also record a published NuGet package putting Opal Chat inside the CMS for editors, so the platform can be embedded as well as called.

Sourcedocs.developers.optimizely.com, Optimizely MCP server documentation and Opal release notesread 2026-09-02

A customer facing evaluation product ships with Opal. Output Evaluation scores specialized agent runs using an LLM-as-Judge model against quality standards the customer defines, with the product page putting it plainly: define what good looks like across tone, accuracy and completeness, and every output is scored against those standards before it reaches a human reviewer. Results are readable and categorical, since Execution Guardrails assign each run a status from Passed, Failed, Warning, Terminated, Rescued, Unapproved or Learning.

Existing specialized agent executions can be added as Output Evals, a regression set built from real runs; Learning mode observes the first twenty successful runs of an agent version to build a baseline before enforcing against it, which is version-aware evaluation; and Anomaly rescue injects corrective guidance into flagged runs. Optimizely's experimentation engine tests experiences, but this separate layer tests the agents themselves.

Sourcesupport.optimizely.com Opal glossary, execution outcomes documentation and optimizely.com/products/airead 2026-09-02

No browser control or computer use is documented. Agents act through registered tools, the MCP server and MCP connectors, native product integrations across the Optimizely suite, and system tools published per product, and one documented external analysis capability, pulling web performance data, runs through an analyze_pagespeed tool rather than by loading and reading a page as a user would. Nothing published describes an agent driving an interface, controlling a browser or operating software that exposes no programmatic route.

Sourcesupport.optimizely.com Opal glossary and release notes, docs.developers.optimizely.comread 2026-09-02