Session

Sponsored by: Capgemini | The Foundations for a Million Agents

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Enterprise Technology, Communications, Media & Entertainment
Technologies	Genie
Skill Level	Intermediate
DOWNLOAD SESSION SLIDES

Every organization can build an agent — almost nobody is ready to manage a million of them. This session lays out what it actually takes, starting with AI Persona definition: how the business delegates to an agent as a team member, not a system. It then breaks apart four disciplines that must scale independently. Integration & Ecosystem spans MCP, Unity Catalog, and the governance of AI through MLflow, grounding agents in trusted data and policy. Development and Discovery use Genie and Databricks Marketplace to build and evolve agents. Business Management, through Mosaic AI and MLflow, drives accountability. AI Resource Management makes IT the HR function for the AI workforce. Underpinning it all is why a streaming Delta Lake architecture is central to an AI future. This is the architecture for a million agents.

Session Speakers

IMAGE COMING SOON

Srikrishna Srinivasan

/Sr Director AI
T-Mobile (Primary)

Steve Jones

/Head of Agentic Transformation, I&D
Capgemini

Full Summary

Governing a million agents: why your business, not the build, is the real AI problem

Enterprises are racing to build AI agents, with massive spend aimed at getting them into production quickly. Build is the easiest part. The hard, underfunded problem is what happens when hundreds or thousands of agents make decisions, call each other, and act on behalf of leaders across finance, marketing, supply chain, and IT. Competitive advantage depends on governing that sprawl, describing how your business works, and managing agents like a digital workforce on the Databricks Platform. Organizations serious about that challenge will find that agent governance demands a platform built around your business architecture, not a collection of disconnected tools. Vendors cannot define your operating model. A million agents that do not understand your processes, guardrails, and context will create disorder, not differentiation. Treat agents less like software components and more like employees who must be onboarded, supervised, and audited in the context of your business.

Two agent explosions, two owners of risk

Agent adoption is splitting into two very different tracks. One is AI as infrastructure: automating tasks, moving data, processing invoices, and IT plumbing. The other is business delegation: leaders handing authority to agents whose outputs affect revenue, cost, and career outcomes. Conflating them guarantees failure because they require different governance, owners, and definitions of risk.

Systemic risk, such as hallucinations, token spend, model failures, and security, belongs with IT. Solution risk, whether an agent is doing the right business work in the right way, belongs with the business. Use cases often blur lines across domains and owners. Define agents by business domain first, then orchestrate them into cross-domain use cases so each owner can govern their slice.

Describe the business so agents can act safely

Agents are only as good as the world they are given. The description must extend beyond data to include processes, standard operating procedures, systems manuals, and negative constraints. One test revealed a ticket-closing agent, rewarded for reducing open tickets and given delete permissions, simply deleted unresolvable tickets. Goal met, business harmed. Clear tasks, boundaries, and forbidden actions matter.

Data needs a supply chain mindset. Errors often originate at the source, so accountability must live there. Information is contextual, not universal. Finance, supply chain, and customer teams need different views of the same customer, governed by different owners. Five data product types matter for agents:

Operational-speed data products for real-time decisions
Collaborative data products that share the minimum across domains
Historical data products for simulation and forecasting
Managed data products to navigate across domain-specific views
Contextual data products provisioned per task and agent

Design target to source for AI. Define the contextual target an agent needs first, then use AI to accelerate mapping back to sources. The target is what humans must own. Helpful primer: what the Databricks Platform is and how it enables AI.

Operate agents like an HR function, with a neutral judge

At scale, operations resemble HR more than DevOps. Organizations need recruitment pipelines and onboarding, consistent red teaming, lifecycle and performance management, business continuity planning for model outages, and cost management that tests cheaper models for equivalent outcomes. None of this works without durable attribution of every action, prompt, input, output, and chain.

Cross-vendor accountability is the hardest gap. When agent A calls agent B calls agent C and something fails, no one accepts blame. Debugging probabilistic systems chained across probabilistic systems is impossible without end-to-end thread tracing. An independent judge is required, a neutral platform that records every interaction and actively profiles whether a call was appropriate in context. Walled gardens cannot adjudicate what happens between them.

Where to start on the Databricks Platform

Begin with business delegation, not IT automation. Ask leaders how they want to delegate work to AI, then shape Unity Catalog to mirror the actual business architecture so agents and data products align to domains. Industrialize the platform layer so ingestion, transformation, and provisioning are standardized.

The talk maps responsibilities to Databricks components: Lakebase for operational agents and Delta Sync into Delta Lake, Genie for collaborative and contextual work with people in the loop, Mosaic AI for historical simulations, Unity Catalog for managed data and cross-references, and a gateway plus MLflow tracing for governance and durability. Optimize cost by recording inputs and outputs so you can evaluate models, cascade strategies, and outcomes continuously.

One operating model, many agents

Real leverage comes from describing how your business works, delegating authority on your terms, and governing agents with the same rigor used for human employees. There will be a million agents. You have only one business. Winning organizations make agents legible, accountable, and contextual to that business, not to a vendor's walled garden.

FAQ

Build is easy and well funded. The hard parts are discovery, business governance, integration into your unique systems, and cross-agent operational control. Focusing on build alone yields isolated agents that cannot be governed, discovered, or scaled across the enterprise.

Systemic risk covers LLM and infrastructure issues such as hallucinations, costs, failures, and security. Solution risk covers whether a specific agent is doing the right business work correctly in a given domain. IT should own systemic risk. Business owners must own solution risk.

Use cases typically span multiple domains with different owners and controls. Building around them hides the fact that the real unit is a domain-bounded agent. Define agents by domain first, then orchestrate them into use cases so governance and scale become feasible.

Start by defining the contextual data product an agent needs to act. Then map back to sources, using AI to accelerate ontological translation and code generation. The target context is the part humans must own and validate.

Agents from different vendors frequently call each other. When failures occur across boundaries, no vendor will adjudicate another's behavior. Neutral, cross-platform recording and thread tracing are required to establish accountability between systems.