Amazon Bedrock models
The Converse API with streaming and tool calling, the same code across model families. Cross-region inference profiles can be limited to one geography, so requests stay within, for example, the EU.
AI Agents · Amazon Web Services
AI agents on AWS.
In your account, under your controls.
Amazon Bedrock gives access to models such as Claude, Llama, Mistral and Amazon Nova inside your AWS account, with your IAM permissions, regions and logging. We design the agent around it: models and guardrails, the workflow that orchestrates steps, the data it can reach and the voice or phone channels it speaks through.
A specialist practice of MLAIA Data ScienceLed by Dr. Yochai Edlitz · Ph.D., Weizmann Institute
The building blocks
Pick the managed pieces.
Write code where the logic is yours.
AWS offers managed agent features and plain compute. The right mix depends on how much of the workflow is standard, how much control you need over each step, and which data and residency rules apply.
Bedrock Agents & AgentCore
Managed agents with action groups, and AgentCore for running your own agent code with managed runtime, memory, identity and tool gateways. Useful when you want AWS to operate the agent runtime.
Knowledge Bases & retrieval
Managed ingestion and vector search over S3 documents, or your own pipeline with OpenSearch or Postgres and pgvector when you need custom chunking, permissions or hybrid search.
Guardrails
Content filters, denied topics, sensitive-information masking and grounding checks applied to inputs and outputs. They complement, not replace, tool permissions and evaluation.
Lambda & Step Functions
Tools as Lambda functions with narrow IAM roles; long or multi-step processes as Step Functions workflows with retries, timeouts and human approval tasks.
Your own agent code
LangGraph or plain Python agents on Lambda, ECS Fargate or AgentCore, calling Bedrock models. Full control over each step, still inside your account. See LangGraph agents.
Voice & contact centre on AWS
The same agent,
speaking on the web or the phone.
AWS covers the speech channel end to end. Amazon Transcribe provides streaming speech-to-text, including Hebrew; Amazon Polly provides text-to-speech; Amazon Nova Sonic offers a single speech-to-speech model for supported languages; Amazon Connect and Lex put an agent on a phone number.
In a healthcare voice assistant we are developing, speech is transcribed by Amazon Transcribe streaming in an EU region and answered by Claude on Bedrock through an EU-only inference profile; the code refuses to start with any other region. The voice agents page covers latency, interruption and voice choice in detail.
Web voice
Browser audio over WebSocket to a streaming STT service, the agent, and a TTS voice streamed back sentence by sentence.
Phone
Amazon Connect routes a call to the agent, with transfer to a human queue and the conversation summary passed along.
Speech-to-speech
Nova Sonic handles listening and speaking in one model, which can reduce delay. We compare it with a pipeline on your languages before choosing.
Services we work withBedrockAgentCoreKnowledge BasesGuardrailsLambdaStep FunctionsECS FargateAPI Gateway WebSocketTranscribePollyNova SonicConnectS3OpenSearchSecrets ManagerCloudWatch
Security & cost
Least privilege for tools.
Hard limits for spend.
Most agent risk on AWS sits in what the tools can touch and how much the agent can spend. We set both explicitly.
Narrow roles per tool
Each tool gets its own IAM role, scoped to the resources it needs. The agent never holds broad credentials, and access keys are not stored on servers.
Region & data residency
Models, transcription and storage pinned to the regions your policy allows, with logging configured so prompts and audio are not kept unless you decide to keep them.
Budgets & throttles
Per-user and per-session rate limits, token and tool-call caps per turn, and AWS Budgets alarms that can disable a public endpoint automatically.
Audit trail
Every tool call recorded with who, what and when, without storing the user's words. CloudTrail covers the AWS side.
A practical delivery path
Start small in your account.
Then harden and scale.
Choose the use case
Define users, channels, permitted actions, data sources and the regions and policies that apply.
Prototype on Bedrock
Compare candidate models on representative tasks; decide between managed agents and your own agent code.
Tools & channels
Implement tools with narrow roles, add retrieval, voice or phone channels, and approval steps.
Harden & operate
Infrastructure as code, budgets, alarms, tracing and an evaluation suite that runs on every change.
Before we start
Good questions. Straight answers.
Does our data leave our AWS account when using Bedrock?
Bedrock runs models within AWS, and you choose the regions. Data handling depends on the model provider and your configuration; we document which services, regions and logs each request touches before go-live.
Bedrock Agents, AgentCore or our own code?
Managed agents are quickest for standard tool-calling assistants. AgentCore suits teams that want their own agent code with a managed runtime. Fully custom code on Lambda or containers gives the most control. We often prototype with one and keep the option to move.
Can the agent answer phone calls?
Yes. Amazon Connect can route calls to the agent through streaming speech services, with transfer to a person when needed. We measure end-to-end delay on real phone audio, which is narrower-band than web audio.
Does it support Hebrew?
Amazon Transcribe supports Hebrew streaming transcription and Bedrock models handle Hebrew text. Speech output in Hebrew usually needs a voice from another provider, which we connect alongside AWS.
How do you control cost for a public-facing agent?
Rate limits per user and session, short sessions, caps on tokens and tool calls per turn, a bot check before expensive steps, and a budget alarm that turns the endpoint off at a fixed monthly amount.
AI Agents & Voice Agents · AI Agents on AWS
Tell us what
you’re working on.
Share the problem, the data you have and what success would look like. We’ll discuss a practical next step.
Please don’t include confidential datasets, credentials or patient information.