AWS показала, как data mesh становится фундаментом для агентного AI в продакшне

AWS published a detailed technical guide on building a managed serverless data mesh — an architecture that provides agentic AI with a reliable, secure, and scalable foundation for production work.

Why agents need data mesh

Agentic AI fundamentally differs from chatbots: it independently forms chains of actions, calls tools, and makes decisions without human involvement. To do this reliably, an agent must receive current corporate data in real time — and do so strictly within security policies. The problem is that corporate data is traditionally scattered across dozens of systems, teams, and formats.

Centralized data warehouses and data lakes solve part of the problem, but create a bottleneck: data becomes stale, schemas diverge, and a single team of engineers cannot handle requests from all departments. Data mesh proposes a different approach: each domain (finance, marketing, logistics) owns its own data and publishes it as a "product" — with documentation, clear schemas, and access contracts. An AI agent connects directly to such a product without intermediaries.

Architecture on AWS

AWS offers a serverless stack that enables data mesh implementation without server management:

• Amazon S3 — decentralized storage: each domain keeps data in separate buckets
• AWS Glue Data Catalog — unified catalog of schemas and metadata for all domains
• AWS Lake Formation — access rights management at table, column, and row levels
• Amazon Athena — serverless SQL queries against data in S3 without additional ETL pipelines
• AWS IAM — role-based model: each agent receives minimum necessary permissions

The key point: agents don't need to know the physical location of data. They access the logical product through the Glue catalog, not directly to S3 buckets. This allows teams to move and restructure data without changing the interface for agents. In practice, the cycle works like this: the agent receives an IAM role, Lake Formation checks permissions on each query, Athena executes SQL, the result returns to the agent. Everything is serverless, all actions are auditable.

Security and control

Agentic AI is unpredictable: it can formulate a query that overloads the system or attempt to read data outside its domain. AWS addresses this at multiple levels. Lake Formation policies block any queries outside the permitted scope — even if the agent technically has access to S3.

CloudTrail maintains a complete audit log of every action. Athena quotas limit the volume of data scanned per query, protecting against accidental or intentional overloads. Another important detail is data semantics.

The Glue Data Catalog stores not just schemas, but field descriptions: the agent understands the meaning of what it reads, not just a set of bytes without context. Schema versioning ensures that structure changes in one domain don't break agents working with that product.

What this means

Data mesh stops being a theoretical concept and becomes a practical requirement for everyone building production AI. Companies without managed data access will get agents that are either unreliable or unsafe — or more often both. AWS offers a ready-made serverless stack that covers most tasks out of the box.