Amazon Bedrock
AgentCore
The platform to build, connect, and optimize agents. Any framework, any model.
Ship agents fast
Build and deploy agents without months of infrastructure work
Connect to anything
Give agents secure access to MCP servers, APIs, and knowledge bases
Optimize continuously
Trace every decision, evaluate performance, safely test improvements
What is an agent?
A computer program that works independently to achieve the goals you give it — and interacts with its environment to get the information it needs.
Everything you need for getting agents to production
Composable by design — click any capability to explore it or jump into its live demo, then compose an agent from just the pieces it needs.
AGENT HARNESS
CONTEXT AND TOOLS
OPTIMIZATION
ENVIRONMENT
SECURITY AND
GOVERNANCE
AgentCore is modular: you turn on only the capabilities your agent needs, and pay only for what you use. Pick an example:
Managed harness: you bring the logic, we handle the infrastructure
Build agents as configuration. Pick options on the left, then run CreateHarness → InvokeHarness. Zero infrastructure config.
🧠 Models Pick your model
🔧 Tools Attach your tools
💾 Memory & Skills Plug in if needed
🪪 Identity & Evals Connect & configure
# Click ① CreateHarness to begin
AgentCore Runtime: every session in its own sealed box
Deploy with one command, from a laptop prototype to production scale. Works with any framework (Strands, LangGraph, CrewAI…), any model, A2A · MCP · HTTP · AG-UI.
👥 Users
A model knows the world, but not your business
Each turn, the agent assembles these context sources into a single LLM prompt.
Give agents a memory that outlasts the conversation
Events land in short-term memory synchronously; extraction runs asynchronously into long-term memory — summaries, preferences, facts and episodes.
📝 Summaries
💖 Preferences
📌 Facts
🎬 Episodes
Each memory is stored as an embedding. At retrieval the query is embedded too, and the closest memories by meaning are returned — not just keyword matches.
Serverless AI gateway: one secure endpoint for tools, agents & models
Gateway as an LLM gateway: one endpoint for every model provider
Point the OpenAI or Anthropic SDK at the Gateway’s /inference endpoint. The model field picks the provider; the Gateway adds auth, policy, guardrails, token limits and the provider’s credentials.
/inferencePick a model and send a request.
AgentCore Browser & Code Interpreter
Secure, isolated execution environments: a headless browser to operate web apps, and a sandbox with shell & file system to execute code.
🎫 New IT ticket
Letting agents pay for the data they need
When an agent needs paid APIs, paid MCP servers or paid data, AgentCore Payments handles the payment through a single API.
How it flows
- User asks: “Analyze Amazon stock”
- Agent needs paid market data
- Resource returns an x402 payment request
- AgentCore Payments handles payment
- Agent gets data to complete the task
- Agent responds to the user
Improving agents with a continuous loop
Observability traces every decision; evaluations score quality; insights & recommendations propose fixes; A/B tests validate them on live traffic.
improvementobserve → evaluate → improve → deploy
💡 Insights · failure analysis PREVIEW
🧪 Recommendation · system prompt
- You are a support assistant.+ You are a support assistant. Before calling+ refund_order, verify the order ID with+ lookup_order and confirm the amount.
⚖️ A/B test on live traffic Gateway routes config bundles per session · online evals score every session
✅ 13 built-in evaluators LLM-as-judge & code-based · or define your own
Temporal policies: deterministic control over autonomous agents
Stateful, trajectory-aware authorization evaluated at the Gateway perimeter — deny-by-default, outside the agent loop.
🤖 Trading agent · session t-01
🙋 Human approval required
📜 Active policies + existing stateless RBAC
// Evaluated policy appears here (Dogwood temporal policy language, Cedar-compatible)
Running agents in production: CI/CD + ongoing operations
Quality gates before release, the AWS Agent Registry for lifecycle, Gateway traffic splitting for canaries, and evals + observability after.
📈 Monitoring & evals Observability · Evaluations · CW alerts
🔀 Gateway traffic split
🗂️ AWS Agent Registry version history
Rollback = the Gateway routes traffic back to a known-good version kept in the Registry. No rebuild or redeploy, so it takes seconds.
Multi-tenant agents: one agent product, many customers
When one agent serves many tenants, every AgentCore component needs a tenancy decision. Explore the concerns, then choose pool or silo for each component.
🧩 Agent-as-a-Service: pool or silo per component
Tenant context and identity: who is calling, for which tenant?
Bind the tenant to the user in the IdP token, validate it on the way in, and pass it on down-scoped on the way out.
🪪 Binding tenant identity to user identity
🤝 Calling tools: impersonation vs act-on-behalf
Tenancy patterns: silo, pool and bridge
Decide how much of the stack each tenant gets to itself. Send a request from each tenant to follow its path.
Tenant isolation at the Gateway: three checks on every call
A tool call to the Gateway passes three checks in order, each finer-grained than the last. This Gateway serves tenant ABC.
Checks who is calling
allow if tenant_id = "ABC"Checks what they may do
refund_order only if amount < 100 and group = "Admin"Checks the tool is enabled for the tenant, cleans data
registry: ABC → lookup_order, refund_order// responseTenant isolation in pooled AgentCore Memory
One Memory resource, many tenants. The runtime turns the tenant in the JWT into scoped AWS credentials, so each tenant can only reach its own actorId and namespace.
🪜 What happens on each call
- Extract tenant_id and sub from the incoming JWT
- Map them to an IAM session tag: actorId = tenant/user
- Call sts:AssumeRole on the ABAC role, with that tag
- Call AgentCore Memory with the tenant-scoped credentials
// API results appear here
sub: userA
📜 ABAC role policy
{ "Effect": "Allow",
"Action": ["bedrock-agentcore:CreateEvent",
"bedrock-agentcore:ListEvents",
"bedrock-agentcore:GetEvent"],
"Resource": "arn:aws:bedrock-agentcore:*:*:memory/pool-memory",
"Condition": { "StringEquals": {
"bedrock-agentcore:actorId":
"${aws:PrincipalTag/actorId}" } } }
🗂️ Namespace levels click one
Observability and cost per tenant
Emit metrics with tenant_id, aggregate them daily, and split each shared service’s cost by usage.
💰 Cost per tenant Amazon Bedrock · total $60
🧮 Sample records
Start building with Amazon Bedrock AgentCore
Any framework, any model. Start with one capability and add more as you grow.
👩💻 Coding agents
Run for hours in persistent, isolated sessions, many in parallel.
⚙️ Workflow agents
Event-driven agents that triage, route and act on emails, tickets and alerts.
💬 Conversational agents
Resolve customer and employee issues faster, connected to CRM, ITSM and knowledge bases.
