Up to 50% Faster Campaign Execution: How the Code Vision Built Virtual CMO with Claude on Amazon Bedrock

Running always-on sales & marketing operations with continuous content production, distribution, competitive research, and campaign execution — is difficult to scale with a lean team without either introducing brand risk (unsupported claims, off-brand messaging, unreviewed publication) or requiring unsustainable manual review effort. The Code Vision partnered with Cygnet to build Virtual CMO: a multi-agent system with Claude as its reasoning engine, research, coordinating content production, and distribution under governed autonomy limits and human review, rather than a single monolithic content-generation tool.

Virtual CMO runs five specialised agents — Orchestrator, Research Intelligence, Content Production, Content Distribution, and Webinar Operations — on Amazon Bedrock, coordinated through the Strands Agents SDK, with Cohere Embed v4 powering retrieval over brand and knowledge-base content. Claude Sonnet 4.5 handles reasoning-heavy content and research tasks; Claude Haiku 4.5 handles lighter-weight, high-volume tasks. The system is governed by application guardrails, instruction-hierarchy controls against prompt injection, and bounded autonomous step limits, each of which has already been exercised in production.

At the core of Virtual CMO is Claude, accessed through Amazon Bedrock, running in production as the reasoning engine behind content generation, competitive research, and campaign distribution decisions.

Model: Claude Sonnet 4.5 on Amazon Bedrock is used for deeper market, competitor and ICP research, campaign strategy, conversion architecture, complex synthesis and high-value content generation. Claude Haiku 4.5 handles request classification and routing, summarisation, extraction, metadata processing, channel adaptation, first-pass variants and supported image captioning. Cohere Embed v4 provides the embeddings used for private knowledge retrieval.

What Claude on Bedrock does today

• Orchestration – validates the user-composed workflow, reserves the permitted credit, determines which nodes are ready, and executes specialist agents in dependency order within governed step, time and cost limits.
• Research Intelligence – performs controlled market, competitor, trend and ICP research through Tavily and SERP API, treating retrieved web content as untrusted context and retaining source information for human verification.
• Content Production – creates researched, brand-aligned marketing content using approved private knowledge where applicable, attaches supporting citations and flags unsupported information for review rather than fabricating missing facts.
• Content Distribution – recommends suitable channels and adapts approved source content for those channels while preserving the approved message and call to action. Publication and send actions remain outside the autonomous baseline and require human-controlled execution.
• Webinar Operations – produces an end-to-end webinar campaign kit by coordinating Pain Miner, Conversion Architect, seven channel-writer stages, Critic, Lead Routing and Exporter. The Critic quality gate permits no more than two automated revision passes before escalation or human review.
• Document Export – supports markdown, docx, xlsx output, with a Markdown fallback path available if structured document rendering cannot be completed successfully.
• Cost Governance – each chargeable workflow is subject to pre-run credit reservation and configured model, tool and workflow budgets, preventing execution beyond the authorized spend envelope. Actual usage is reconciled when the workflow completes.

Responsible AI

Claude’s inputs and outputs in Virtual CMO run behind layered governance: instruction-hierarchy defenses treat all retrieved research and user-submitted content as untrusted data rather than instructions, so an embedded directive in a competitor’s web page or a knowledge-base document cannot redirect agent behavior. Application-level guardrails hold regulated-domain or otherwise unsupported marketing claims at draft status pending specialist review before anything is published. Bounded autonomous step limits prevent a broad or poorly-scoped research brief from running away. An indirect prompt-injection attempt embedded in retrieved third-party web content was rejected without altering agent behaviour. Each of these controls has already been exercised against a real production event and performed as designed.

The Code Vision partnered with Cygnet to build Virtual CMO on AWS, with Claude on Amazon Bedrock as the reasoning engine behind the platform’s content and campaign operations. The engagement began with a discovery phase covering the marketing workflows to be automated, the brand and knowledge sources content should be grounded in, the external research dependencies required, and the governance controls needed before any agent could publish or distribute content on the customer’s behalf.

Cygnet designed Virtual CMO as five coordinated agents rather than a single generation tool. An Orchestrator agent interprets each content or campaign request and sequences the work across Research Intelligence, Content Production, Content Distribution, and Webinar Operations agents, built on the Strands Agents SDK. Claude Sonnet 4.5 performs the reasoning-heavy work — drafting content, synthesising research, and making distribution decisions — while Claude Haiku 4.5 is routed lighter-weight, high-volume tasks, both accessed through Amazon Bedrock.

Content grounding was a specific design priority: rather than generating marketing copy from the model’s general knowledge, the Content Production agent retrieves relevant brand and knowledge-base context through Cohere Embed v4 before drafting, reducing the risk of off-brand or unsupported claims. Where current external context is required, the Research Intelligence agent performs live web research through Tavily and SERP API, with citation and currency checks applied to retrieved content before it is incorporated into a deliverable.

Because retrieved web content and user-submitted briefs cannot be fully trusted, Cygnet applied instruction-hierarchy controls that treat all such content as untrusted data rather than executable instructions — a control that has already contained a real indirect prompt-injection attempt embedded in a competitor web page during a Content Distribution run. Application-level guardrails separately hold regulated-domain or otherwise unsupported marketing claims at draft status pending specialist review, and bounded autonomous step budgets stop an open-ended or poorly-scoped research brief before it runs away — both of which have also been exercised against real production events and performed as designed.

Deliverables are exported to DOCX or XLSX for the customer’s existing workflows, with an automatic fallback to Markdown if a structured export fails, so a formatting issue never blocks delivery of the underlying content.

Cost governance is enforced through credit-reservation controls and per-workflow budgets across Amazon Bedrock inference, retrieval, and research API usage, so no workflow can exceed its approved spend before execution.

Having established a governed, production content pipeline, the Code Vision is now turning to campaign-effectiveness measurement tracking how Claude on Amazon Bedrock-driven research, content, and distribution decisions translate into engagement and pipeline outcomes, to guide the next round of workflow and model-routing refinement.