Amazon Bedrock AgentCore · delivered by CloudZA®

Production AI agents, built to last.

AgentCore gives agents a secure runtime, memory, identity and tools. CloudZA® designs, ships and operates them on AWS — from a first proof-of-concept to an agent your business depends on, then we stay for support.

Runtime to production Secure by design Observable & supported
$ cloudza agentcore deploy --agent support-copilot
 Runtime        provisioned · scale-to-zero
 Memory         long-term store linked
 Identity       least-privilege roles bound
 Gateway        6 tools exposed via MCP
 Observability  traces streaming
 agent live: prod  ·  p95 1.2s
The capability map

Seven building blocks. One production agent.

AgentCore is modular — use what you need. Here's each piece, and how CloudZA® puts it to work.

Build & run

Runtime

A serverless runtime to deploy agents in any framework or model and scale them to zero.

CloudZA®: we package your agent and ship it to prod, no servers to run.
Remember

Memory

Short- and long-term memory so agents recall context across turns and sessions.

CloudZA®: we design exactly what the agent should remember — and forget.
Secure

Identity

Secure agent identity and least-privilege access to your tools and data.

CloudZA®: every agent gets scoped credentials — never a standing key.
Act

Gateway

Turn your APIs and Lambda functions into agent-ready tools via MCP.

CloudZA®: we expose your systems to the agent safely, one tool at a time.
Act

Browser

A managed, isolated browser so agents can navigate and act on the web.

CloudZA®: agents that complete real web tasks, sandboxed and logged.
Act

Code Interpreter

A sandboxed environment for agents to write and run code on the fly.

CloudZA®: agents that compute and analyse — not just chat.
Observe

Observability

End-to-end tracing and evals of every step an agent takes.

CloudZA®: we watch quality, cost and drift — and fix it before you feel it.
Foundation models

Bedrock models

Frontier and open models through Amazon Bedrock, governed in your account.

CloudZA®: we pick the right model per task and keep your data yours.
The agent harness

An agent is a configuration on a managed harness.

AgentCore separates what your agent is from what runs it. CloudZA® owns both — so the same agent behaves the same from a laptop experiment to production.

Agent configuration · what it is

The declarative spec

Versioned and reviewed like code — model, prompt, tools and the guardrails on the loop.

LLM modelSystem promptToolsSkillsAllowed toolsMemoryMax iterationsMax tokensTimeout
Managed harness · what runs it

The AgentCore primitives

The managed services CloudZA® wires in and operates — no agent infrastructure to run yourself.

RuntimeMemoryIdentityGateway (MCP)BrowserCode InterpreterObservabilityEvaluationsStrands agents
Why a managed harness

The hard parts of agents, handled.

Teams stall when an agent meets reality — sessions, isolation, tenancy, tool selection. The harness takes that on so we ship.

Rapid prototyping

Stand an agent up in minutes — two parameters to start, everything else layered on later.

Personalization

Memory and per-user context so the agent adapts to each person it serves.

Multi-tenancy

Isolated sessions and scoped identity keep every tenant's data and actions cleanly apart.

Context & tool selection

The harness manages the agent loop — choosing the right tools and trimming context for you.

Coding agents

A sandboxed environment with shell and code execution, for agents that build — not just chat.

Long-running agents

Conversational and long-running workloads with sessions that live for hours, then clean themselves up.

Build · run · operate

From two parameters to production.

A harness needs only a name and an execution role to exist. CloudZA® layers on model, prompt, tools, memory and guardrails — as code, reviewed and repeatable.

cloudza/create_harness.py
# CloudZA® provisions the harness as code
client = boto3.client("bedrock-agentcore-control")

harness = client.create_harness(
  harness_name="support-copilot",
  executionRoleArn=role_arn,            # only 2 required
  model={"bedrockModelConfig": {"modelId": "claude-sonnet"}},
  systemPrompt={"text": "You are CloudZA®'s agent…"},
  memory=memory_cfg,
  tools=[gateway_tools],
  skills={"path": skills_path},
)
model · systemPromptWhich Bedrock model the loop uses, and the agent's instructions.
tools · skills · allowedToolsGateways, browser, code and reusable skills — scoped to what the agent may call.
memoryShort- and long-term recall configuration.
authorizerConfigurationAgent identity and least-privilege access.
maxIterations · maxTokens · timeoutSecondsGuardrails on the agent loop and its cost.
environment · environmentArtifactRuntime config, or bring your own container image.

Runtime & session lifecycle

Every session runs in an isolated microVM with a managed lifecycle — so we never leave compute, or cost, running.

microVM

Hard isolation per session — the default image, or bring your own container from ECR (linux/arm64).

idle 15m

Idle timeout reclaims a session after inactivity, so nothing lingers.

max 8h

Max lifetime caps long-running sessions; network and filesystem are configurable.

shell

Direct shell access to the harness VM for pre/post scripts, environment setup or installing skills — not everything has to go through the agent loop.

observe

Observability & evaluations trace every step and score quality before a change ships.

How we deliver · POC → PROD → support

Most agent demos never ship. Ours run.

01 · PROVE

A real PoC in weeks

One use-case, your data, measured against a business metric — built on AgentCore Runtime and Bedrock.

02 · PRODUCTIONISE

Hardened for prod

Identity, memory, gateways, evals, guardrails and cost controls — Well-Architected, as code.

03 · OPERATE

Watched & evolved

We run it with full observability, tune quality and cost, and ship the next capability — long term.

Start the conversation

Let's prove an AgentCore use-case on AWS.

Run the readiness assessment, then a CloudZA® architect scopes an agent against a metric that matters — and takes it from PoC to production.