Marketing Technology · Enterprise AI

Optimizely launches role-based Virtual Teammates and announces post-trained marketing models for its Agent Platform, with a benchmark it describes as open source.

By Shashi Bellamkonda · October 25, 2026

Role-based Virtual Teammates arrived from Optimizely on August 31, and the announcement of post-trained marketing models followed on September 1, both built for the company's Agent Platform. Together they point the platform toward a model layer tuned to marketing and a staffing layer that assigns work to named roles.

I first met Optimizely as a case study, back when a developer could run an A/B test with nothing more than a credit card. What the company sells today is much larger, with content management, experimentation, commerce, personalization, and analytics on one platform, according to its own description.

When I compare vendors, I have often been in the dark about Optimizely and could research what was public and nothing else, so I appreciate the update.

Post-training turns a general model into a marketing specialist

Optimizely says it removes the excess from state-of-the-art models, leaving versions built for marketing tasks. In early testing the company reports 10x greater cost efficiency than state-of-the-art large language models (Optimizely Models, 2026). Imran Yousuf, Head of AI Engineering at Optimizely, calls post-training the long-term answer to the cost versus quality debate.

The models draw on Mark-IQ, the Agent Platform's data layer, which holds each customer's experimentation history and web analytics and gives every model the brand's context. For a company that began with A/B testing, that history is a distinct asset: it records which version of a page won with real visitors, a different kind of context from the text a general model learned from.

Optimizely says the platform applies the right model for each use case, so marketers do not have to pick one.

Mark-Bench gives you a way to check the 67% on your own marketing tasks

Optimizely also introduced Mark-Bench, a benchmark it describes as open source, covering 285 tasks across 15 marketing functions and more than 6,000 criteria, including writing a press release and drafting email copy (Optimizely Models, 2026). On the benchmark's all-pass rate with default configurations, Agent Platform scored 67% against 60% for Claude Code, at 2x lower cost (Optimizely Models, 2026).

Optimizely wrote the benchmark and published the first score, so the invitation to outside testing is the part to use. In its Mark-Bench introduction, the company says it is open-sourcing the tasks, rubrics, and scoring, and it invites marketers, researchers, and other providers to test their own models and agent harnesses against it. Shafqat Islam, the company's President, argues that specialized harnesses help customers escape pilot purgatory. Ask for the repository and license, then run your ten most frequent tasks to see whether a seven-point gap holds on your own work.

Experiment history records which version won, a different kind of context from the text a general model read.

Virtual Teammates work under their own identity and an audit trail

Each Virtual Teammate has a defined role and organizational permissions, and can run on a schedule or an event trigger. Launch roles include a Chief of Staff, a Marketing Analyst, a Personalization Strategist, and a conversion rate optimization manager. Teams hire them from a directory and adjust their responsibilities in plain language (Optimizely Virtual Teammates, 2026).

Companies set where human review is required, and each teammate's activity traces back to its own identity. Kevin Li, SVP of Product at Optimizely, said teams often cannot tell who did what or trust the AI enough to let it act, which is why the teammates carry an audit trail from day one. Optimizely's own study of more than 2,000 business-to-business marketing leaders found that 81% switch between two or more disconnected AI tools each week (Optimizely Virtual Teammates, 2026), and the teammates are designed to absorb that switching.

Adobe approaches the same problem from brand governance

Adobe coordinates marketing agents through Adobe Experience Platform Agent Orchestrator, which spans planning and operations, audiences, content, journeys, experience delivery, and data insights.

In its April 20 Adobe Summit release, the company added three Adobe Experience Manager agents that were available at announcement. Two of them handle content: the Brand Experience Agent updates and creates pages, and the Content Advisor Agent finds approved content for each channel. The third, the Brand Governance Agent, enforces brand policies, tracks asset rights, and manages permissions.

Adobe frames the work as a four-stage cycle of sense, generate, reach, and learn, aimed at how a brand appears in AI-driven discovery. Optimizely starts from a different asset, its experimentation history, and ties each teammate's activity to an identity. Adobe describes agent work that can be checked and repeated, so ask each vendor to show you the log from one agent action.

Both vendors mix shipped and announced items, so ask each for a dated roadmap.

Four more vendors attach agents to different layers of the stack

The rest of the field sells agents through a layer you may already pay for, and that layer decides which contract you negotiate.

• Salesforce: It completed its purchase of Contentful on September 1, which gives Agentforce a native content layer. At Dreamforce 2026, Salesforce announced job-ready agents, long-horizon agents with approval checkpoints, and Koa, a reasoning model it trains on synthetic scenarios and runs in its own infrastructure.
• Braze: At its Forge 2026 event, Braze released four beta offerings, including Agentic Standards, which checks campaigns against marketer-defined rules and logs each check and fix for audit. CMSWire reported that Salesforce announced on October 20 an agreement to acquire Braze for about $5.66 billion in enterprise value, with closing expected in Salesforce's fiscal 2028, which begins in February 2027 (CMSWire, 2026).
• HubSpot: Agent Hub is live, and the Campaign Agent and Nurture Agent are in public beta. MarTech reported published unit costs of about $0.10 per nurture email and about $10 per Content Agent asset (MarTech, 2026).
• Sitecore: SitecoreAI, announced November 4, 2025, launched with an Agentic Studio holding 20 agents and was available at announcement (Sitecore, 2025).

Each vendor sells agents through a layer you may already pay for: content and, once the deal closes, engagement at Salesforce, brand governance at Adobe, and experimentation at Optimizely. The layer you can least afford to leave should drive your selection.

HubSpot publishes unit prices for its agents, so ask the others for the same.

Two open items will shape the buying decision

Availability comes first. At the September 1 announcement, Optimizely said the models were coming soon to all Agent Platform customers, and an October 5 post on Optimizely World still lists them as announced, not shipped. The same post says Virtual Teammates are not yet available to every customer, partners in particular. Ask your account team for the model date, the pricing basis, and your own access, because a cost-efficiency claim helps your budget once you know how model usage is billed.

Identity is second. Because each teammate acts under its own identity, your security team will want to see how those identities connect to your existing access management and where the activity log lives.