Reference

The Agentic Stack

The building blocks of an agentic AI platform, each mapped three ways: the capability, the popular open-source tools, and the AWS-native service that does the same job. One shared map under everything we build, whether you are building software with agents or shipping an agent to your own customers.

Two jobs, one stack

The same building blocks, weighted differently

People mix these up, so we name them. The stack is the same for both. What changes is which parts carry the most weight.

Building with agents

Using agents to build software faster. The org and process side leads here: who owns the platform, and who helps the teams adopt it.

Running agentic systems

The product itself is an agent your customers use. Identity, sandboxing, guardrails, and evals lead here, because the agent is now in front of users.

The map

Capability, open source, AWS-native

AWS-native is our default. Where open source earns its place for portability or control, we say so. The capability is what matters; the tools are just how you get it.

LLM gateway and routing

One front door to every model, with routing, fallback, and cost control.

Popular open source
Agent runtime and harness

Where agents actually run, with their tools and context around them.

Agentic memory

Short and long-term memory, so agents keep context across steps and sessions.

Embeddings, vector and RAG

Grounding agents in your own knowledge and documents.

Popular open source
Tool and context protocol

How agents call tools and talk to each other.

Popular open source
Sandbox and code execution

Running agent actions and generated code safely, isolated per session.

Weighs heavier when you ship agents to customers
Popular open source
Identity and access for agents

Giving agents their own identity and least-privilege access.

Weighs heavier when you ship agents to customers
Popular open source
Observability, tracing and evals

Seeing what agents did, and proving they did it well.

AWS-native

OpenTelemetry-based, so the same traces export to your own tooling too.

Guardrails and policy

Keeping agent output safe, on policy, and free of sensitive data.

Weighs heavier when you ship agents to customers
Popular open source
CI/CD contract enforcement

Eval and contract gates that stop a bad agent change before it ships.

Popular open source
Standard CD pipelines
AWS-native

AgentCore Evaluations scores agent trajectories and outputs; Bedrock Evaluations covers the model and RAG. The managed parts exist; wiring them into a release gate is where most teams still want a hand.

Amazon Bedrock AgentCore has been generally available since October 2025 (Runtime, Memory, Gateway, Identity, Browser, Code Interpreter, and Observability), with Policy and Evaluations reaching GA in March 2026. It is modular and maps almost one to one onto the open-source rows above, which is the whole point of this page: the same capabilities, with a clear AWS-native answer for each.

Where this fits

Where do you sit on this stack?

The map is the easy part. Knowing where you stand, what to build first, and what good looks like for your team is the work we do together.

Not sure yet? Take the free 2-minute AI Readiness Check first. Take the check →