One-page internal reference combining the multi-agent orchestration proof review with the current draft of client-facing capability language.
Two RFPs plus a shared security annex. The core build ask (Enterprise Agentic Framework Foundation) is an infrastructure and platform RFP: stand up Amazon Bedrock as a governed production runtime with a Terraform module library, hardened CI/CD, a security/IAM baseline, observability, and three reference agents: RAG agent, action-taking agent, and a Multi-Agent Supervisor, promoted through Dev, Stage, and Production. A second RFP (Enterprise Agentic Workspace) asks for a governed chat/storefront layer on top of Bedrock and other agent platforms.
The multi-agent requirement is narrow: Section 8.3 calls for a supervisor delegating to documented collaborator agents, deployment serialization, an enterprise-approved memory/persistence approach, and inherited guardrails/logging/IAM across the agent graph. No energy-specific use case (predictive maintenance, load forecasting, etc.) is named anywhere in the RFP.
| Proof point | What it actually is | Status |
|---|---|---|
| ICA 2.0 + watsonx Orchestrate |
IBM's internal multi-agent delivery platform: model routing, domain grounding, prebuilt role-based agents, orchestration engine. | Confirmed Internal IBM platform, not a client build. No metrics, no energy use case. |
| UFC Insights Engine |
IBM Build Engineering + UFC. watsonx Orchestrate routes a pre-fight insights agent to text-to-SQL and RAG pipelines (IBM Granite, Llama). watsonx.governance for explainability, watsonx.data for vectorized retrieval, Terraform for IaC deployment. | Confirmed Client-facing, named, with metrics: 3x estimated increase in insight volume, 40% estimated reduction in query generation time (both labeled "estimated" by IBM, illustrative-only disclaimer applies). Strongest analog to AEP's Multi-Agent Supervisor pattern. IBM Consulting delivered, not Hakkoda. |
| CEMIG | IBM Consulting + CEMIG (Brazilian energy utility, 9.35M customers). watsonx.ai and watsonx Assistant powered virtual assistants with Llama/Hugging Face LLM integration and NeuralSeek. Unified WhatsApp, mobile app, and web portal into one governed customer experience. | Confirmed Named energy-sector client with hard metrics: 30% OPEX decrease, 20-point NPS increase, 90% reduction in claims-related financial compensation. Illustrative-only disclaimer applies. Confirmed agentic AI: watsonx.ai orchestrating virtual assistants across multiple digital channels at scale inside a regulated energy utility. IBM Consulting delivered, not Hakkoda. |
| AutoNation | Named on an internal list as a possible multi-agent / Cortex reference. | Unconfirmed Zero documentation anywhere in project knowledge. Not usable as proof. |
| Beth Israel Lahey Health |
Snowflake + AWS migration. Dynamic ETL via Matillion. 1,000+ Clarity / 400+ Caboodle tables ingested. | Confirmed Real, AWS platform-engineering proof. Zero prior AWS experience, 3-month engagement. Not agentic. |
| Allergan (AbbVie) |
AWS serverless framework migrated to dbt Cloud / Data Mesh. GitHub CI/CD, DataDog/PagerDuty observability. | Confirmed Real, AWS platform-engineering proof. Zero-downtime cutover, 10+ pipelines. Not agentic. |
| Medtronic | AI-powered supplier intelligence app on Snowflake Cortex Analyst. 3 connected semantic models, conversational interface. | Directional Closest analog to RFP's RAG agent pattern. 8-week engagement. No hard metric, "strong accuracy, high user trust" language only. |
| IBM IBV research reports |
Generic multi-agent workflow frameworks, 2027 executive benefit projections. | Methodology only Thought leadership, not a Hakkoda/IBM build. Never cite as proof of a build. |
We have confirmed, production agentic orchestration proof. The UFC Insights Engine is a live, client-facing multi-agent build: watsonx Orchestrate coordinating a pre-fight insights agent across text-to-SQL and RAG pipelines, with watsonx.governance providing explainability end to end — 3x insight volume, 40% reduction in query generation time. CEMIG is confirmed agentic AI deployed inside a regulated energy utility at scale: watsonx.ai orchestrating virtual assistants across WhatsApp, mobile, and web with hard outcomes (30% OPEX reduction, 20-point NPS lift, 90% fewer claims). Both are IBM Consulting builds. The response leads with these confirmed builds first, backed by platform-engineering discipline on AWS (Beth Israel Lahey, Allergan) and grounded-AI discipline on Snowflake (Medtronic). The infrastructure changes across these proof points; the operating standard does not.
We build governed AI and data platforms across AWS, Snowflake, and other modern data platforms — and we have already built the exact architecture your RFP describes: a coordinating agent layer directing specialist agents toward a shared outcome, governed, observable, and secure by default.
Your RFP asks for a Multi-Agent Supervisor pattern: an orchestrator delegating to documented specialist agents, with persistent memory and guardrails that travel with every agent in the graph. We have delivered this in production.
The UFC Insights Engine is a live multi-agent deployment built on watsonx Orchestrate: a pre-fight insights agent coordinating text-to-SQL pipelines and retrieval-augmented generation pipelines built on IBM Granite and Llama, with watsonx.governance providing explainability and quality scoring on every output. The result: an estimated 3x increase in insight volume and an estimated 40% reduction in query generation time. This is the exact pattern your RFP describes — a supervisor routing tasks to specialist agents, governed end to end. IBM Consulting Advantage 2.0 runs the same orchestration engine at enterprise scale across IBM's global consulting workforce, coordinating specialist agents across multiple foundation models and domain-specific knowledge bases, in production.
At CEMIG, one of Brazil's largest power utilities serving over 9 million customers, IBM Consulting deployed AI-powered virtual assistants on watsonx.ai — orchestrating intelligent agents across WhatsApp, mobile, and web into a single governed customer experience at utility scale. The results: a 30% decrease in operational expenses, a 20-point increase in Net Promoter Score, and a 90% reduction in claims-related financial compensation. This is the same engineering discipline your RFP demands, applied inside a regulated energy environment, with outcomes a board can point to.
We've proven this operating model on AWS specifically. At Beth Israel Lahey Health, we stood up a centralized cloud data platform on Snowflake and AWS, with dynamic ETL pipelines ingesting over 1,400 tables at scale, for an organization with zero prior AWS experience, delivered in three months. At Allergan, an AbbVie company, we migrated a legacy AWS serverless framework to a governed Data Mesh architecture, with GitHub-based CI/CD, real-time observability through DataDog and PagerDuty, and zero-downtime cutovers across more than ten production pipelines.
At Medtronic, we built an AI-powered supplier intelligence application on Snowflake Cortex Analyst — three semantic models unified into a single conversational interface, built to enterprise security and architectural standards in eight weeks. The same pattern your RFP names for a Retrieval-Augmented Knowledge Agent: grounded in real enterprise data, producing trustworthy answers, governed from day one.
A Multi-Agent Supervisor on Bedrock is not a new pattern for this team. We bring confirmed agentic orchestration at scale (UFC Insights Engine, ICA 2.0), confirmed deployment inside a regulated energy utility (CEMIG), and platform-engineering discipline across AWS and Snowflake (Beth Israel Lahey, Allergan, Medtronic). The infrastructure changes. The standard does not. Every agent in the graph inherits guardrails and IAM scoping the same way every pipeline at Allergan inherited CI/CD enforcement, and the same way every Medtronic query is grounded in a governed semantic model rather than a black box.