Session

AstraZeneca's Multi-Agent System: Lessons Scaling Agents by 10x With Agent Bricks

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Healthcare & Life Sciences
Technologies	Genie
Skill Level	Intermediate
DOWNLOAD SESSION SLIDES

AstraZeneca built a multi-agent system using Agent Bricks to transform how commercial teams access pharmaceutical data. The Brand Assistant uses a supervisor agent coordinating specialized sub-agents across therapeutic areas, combining structured data via Genie Spaces with unstructured documents via Knowledge Assistant.

Unity Catalog integration with Entra ID enforces strict permission boundaries. MCP unlocks third-party tool integration. The architecture scaled from 5-agent POC to 20+ agents in production, designed for 50+.

Key components:

Multi-Agent Supervisor for orchestration
Genie Spaces for NL-to-SQL
Knowledge Assistant for documents
Unity Catalog for row/column-level security
MLflow for tracing.

This session covers:

Multi-agent architecture patterns
Permission-preserving design
When to split supervisors vs. add agents
Human-in-the-loop testing
Why bad data breaks agents regardless of tech
Organizational challenges of agent ownership
Lessons from scaling

Session Speakers

Brian Burke

/Senior Director, Platform Engineering
AstraZeneca

Homayoon Moradi

/Staff AI Engineer
Databricks

Full Summary

Scaling AI agents at AstraZeneca with a federated, governed platform

Large enterprises are awash in AI assistants that are easy to spin up but hard to govern. AstraZeneca faced the same pattern, with business teams creating Genie spaces, Microsoft Copilot add-ons, and other agents faster than central IT could evaluate or productionize them. The company and Databricks engineered a federated agentic platform on the Databricks Platform that lets domains own their agents while a shared supervisor layer enforces identity, governance, quality, and shared capabilities. The result balances speed and adoption with trust and operational rigor.

A federated model that avoids both extremes

Enterprise AI often fails at the edges, either as a sprawl of unaligned tools with no single source of truth, or as a monolith that business users ignore. Federation splits responsibilities: domain teams build the agents they know how to use and maintain, and a central Agent Bricks supervisor routes queries, applies common policies, and provides shared tools.

A single enterprise question, such as how a specific brand's market share is trending in a given region, can be routed by the supervisor to the right domain Genie space or Knowledge Assistant. Common functions, such as web search and MCP server integrations for regulated, branded content, run centrally so every domain benefits without duplicating work.

Identity-aware trust that survives every agent hop

Trust hinged on proving that user-specific access controls persist across prompts, tools, and data retrieval. Service-principal authentication made the platform blind to each caller, forcing lowest-common-denominator access. An on-behalf-of-user model fixed this by carrying the user's identity through every step, so Unity Catalog row-level security and column masking apply automatically.

Authentication begins at the application, flows into the supervisor, and is enforced down to subagents and data. Microsoft Entra groups manage who can use which agents, and domain owners approve access to their own groups, keeping IT out of one-off entitlement decisions.

Portable visualization that drives adoption

Tables alone did not win executive attention, and static chart images failed to travel cleanly across APIs, Teams, mobile, or custom UIs. A lightweight alternative uses a Unity Catalog function to ask the agent for a Vega-Lite JSON spec that describes the desired chart. Any receiving UI renders it natively and interactively, keeping payloads small and visuals high quality across surfaces.

The approach is agent-first and surface-agnostic. Agents output a compact "recipe," and UIs render it in context, which proved materially better for adoption than shipping screenshots.

Continuous evaluation to keep agents accurate

Agent quality drifts as data, questions, and KPIs evolve. Manual tuning across dozens, soon hundreds, of agents does not scale. AstraZeneca open-sourced MaxGenie, an optimization loop that treats a Genie space's instructions, SQL examples, and column metadata as modifiable code.

Benchmarks are split into train, validation, and holdout sets to avoid overfitting. The loop proposes edits, tests on unseen questions, and iterates until accuracy plateaus. In a demo, a New York taxi Genie improved from 10 percent to 50 percent accuracy in roughly 12 minutes.

Running and governing the platform in production

Unity Catalog provides end-to-end control, from the supervisor's permissions to table, row, and column access. MLflow traces deliver observability for agent runs, including caller identity, inputs and outputs, timing, and execution graphs. Thumbs up or down feedback from the UI lands in a Delta table keyed to traces, enabling triage and iterative improvement.

Operational discipline matters. CI and publishing processes are encouraged, even as some business teams prefer ad hoc "v1, v2" iteration. Living on public and private previews accelerates capability but carries risk. Current Agent Bricks limits include 20 subagents per supervisor and 50 MB per Knowledge Assistant file, while the number of files in a Knowledge Assistant volume is effectively unlimited. Limits are expected to rise, so teams should monitor release notes.

Federation that the business will use and IT can trust

Agent creation is easy, but sustainable scale requires federation with guardrails. AstraZeneca's approach combines domain-owned agents, a central supervisor, identity that follows every call, surface-agnostic visualization, and continuous evaluation. Organizations pursuing similar architectures can apply the same patterns to balance speed with governance.

FAQ

Service-principal authentication hides the caller's identity, so access must be set to a broad common scope. On-behalf-of-user authentication carries the individual's identity through prompts, tool calls, and data retrieval, allowing Unity Catalog row-level security and column masking to apply per user without the agent making access decisions.

Images do not render consistently across Teams, mobile, and custom apps, and they cannot be manipulated. A Vega-Lite spec is a compact chart "recipe" that any UI can render natively, producing interactive, high-quality visuals with minimal payloads.

Benchmarks are split into train, validation, and holdout test sets. The optimization loop proposes changes to instructions, examples, and metadata, validates improvements on unseen questions, and stops when gains plateau, preventing inflated scores that collapse on new prompts.

At the time of the talk, a supervisor could connect to 20 subagents. Knowledge Assistant supported files up to 50 MB, and the number of files in a Knowledge Assistant volume was effectively unlimited. The speakers expected these limits to rise over time.

Access is managed through Microsoft Entra groups integrated with Unity Catalog. Business owners of each agent approve membership to their groups, keeping entitlement decisions with the people who understand the domain.