Automated Workflows and AI Agents with Amazon Bedrock

Amazon Bedrock is a fully managed generative AI platform from AWS, providing secure access to top foundation models, automated agent creation, and seamless integration with enterprise data through its Knowledge Base feature. Security for generative AI revolves around strong access controls, data privacy protection, governance, monitoring, and special defense against unique threats like prompt injection and model poisoning. Amazon Bedrock Knowledge Bases allow retrieval-augmented generation (RAG) workflows and secure model access to proprietary data, enhancing answers while managing privacy.

Amazon Bedrock Overview
Amazon Bedrock acts as a unified API for accessing high-performing foundation models such as Anthropic’s Claude, AI21 Labs, Cohere, Stability AI, and AWS’s Titan models.

It is serverless, eliminating infrastructure worries and automating scaling.

Supports fine-tuning and retrieval-augmented generation (RAG) to customize model outputs for enterprise use cases, including deploying agents that execute tasks using company data.

Provides developer-friendly interfaces and integrates with popular AWS tools like Lambda and SageMaker.

Includes built-in guardrails for controlling input/output and minimizing risks of bias and toxicity, aligning models to organizational standards.

Ensures privacy—organizational data is not used for general model training and stays within the account boundaries.

Security Considerations for Generative AI
Generative AI introduces specific vulnerabilities such as data leakage, model poisoning, adversarial attacks, and prompt injection.

Key mitigation strategies include:

Access control: Use role-based restrictions and multi-factor authentication.

Data protection: Encrypt data at rest and in transit, sanitize inputs, and apply privacy techniques like differential privacy.

Governance and compliance: Adhere to regulations (e.g., GDPR), enforce data handling standards, and maintain audit trails.

Continuous monitoring: Employ tools for anomaly detection, system logging, and vulnerability management.

Adversarial testing: Regularly simulate attacks and patch vulnerabilities.

Explainable AI: Enhance transparency to identify bias and security weaknesses.

Training: Promote awareness around ethical and security concerns in team members.

Amazon Bedrock Knowledge Bases
Knowledge Bases provide a fully managed RAG workflow that augments model prompts with context drawn from an organization's private and structured data sources, enabling relevant and accurate responses.

1. Automated Workflows and AI
Agents with Amazon Bedrock
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
Alexander
Barge
Sr. Solutions
Architect AWS

2. Basic Foundation Model Flow
Inpu
t
Outpu
t
Foundatio
n Model
“a concept car that
flies and can be used as
a boat”
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

3. Workflow automation challenges
LLM’s are
powerful, but
they can’t take
actions
Integration of data and
internal systems is
crucial but challenging
Orchestration of diverse
programming
languages and
interfaces
Enterprise-grade
solutions involve
complex engineering
4
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

4. Security considerations for generative
AI
The policies,
procedures, and
reporting needed to
empower the business
while minimizing risk
---
Create generative AI
usage guidelines
Establish process
for output
validation
Develop monitoring
& reporting
processes
The specific regulatory,
legal, and privacy
requirements for using
or creating generative
AI solutions.
---
Retain control of
your data
Encrypt data in
transit and at
rest
Support
regulatory
standards
The implementation of
security controls that
are used to mitigate
risk.
---
Human-in-the-
loop
Explainability &
auditability
Testing
strategy
Identity and
access
management
Identification of
potential threats to
generative AI solutions
and recommended
mitigations.
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
---
Threat
modeling
Third-party
risk
assessments
Ownership of data,
including prompts
and responses
How to architect
generative AI solutions
to maintain availability
and meet business
SLAs.
---
Data
managemen
t strategy
Availabili
ty
High Availability
and Disaster
Recovery strategy

5. Use as is or
customize
Fine-tune FMs as
needed;
Bedrock will
automatically
deploy the FM
for inference
Send
prompt
Use the
Bedrock API to
send your
prompts to the
model
Choose an
FM
Use the
playground to
experiment
with FMs and
select the one
that suits your
needs
Receive
response
Receive the
model
response in
your
application
Amazon
Bedrock
Build generative AI
applications using FMs
through a serverless
API service
6
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
Fully managed, enterprise-grade access to
FMs

6. Helps keep your
data secure and
private
None of the customer’s data is
used to train the underlying model
All data is encrypted in transit and at
rest
Data remains in the Region where
the API is processed
Guardrails can be customized to
application requirements and
responsible AI policies
Support for GDPR, SOC, ISO,
CSA compliance, and HIPAA
eligibility
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

7. Your data is the
differentiator
Generic
generative
AI
Generative AI that
knows your
business and your
customers
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

8. Common approaches
for
CO M P L E X I T Y ,
QU A L I T Y , CO
S T , T I M E
Prompt
engineerin
g
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
Retrieval
Augmente
d
Generation
(RAG)
Fine-tuning
Continued
pretraining

9. Amazon Bedrock Knowledge
Bases
NA T I V E S U PPO R T F O R R ET R I EV A L A U G M ENT ED G ENER A T I O N ( R A G )
10
MODEL
Anthropic—
Claude
Meta—
Llama
A M A Z O N B E D R O C K
Amazon—Titan Text
AI21 Labs—
Jurassic2
USER QUERY AUGMENTED PROMPT ANSWER
KNOWLEDGE BASES
FOR AMAZON
BEDROCK
32
1 4 5 6
Securely connect
FMs to data sources
for RAG to deliver
more relevant
responses
Fully managed
RAG workflow
including
ingestion,
retrieval, and
augmentation
Built-in session
context
management for
multi-turn
conversations
Automatic citations
with retrievals to
improve
transparency
Mistral AI - Mistral
Cohere –
Command
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

10. From LLMs to Agentic Workflows
Observe results
Think about next
step
Inpu
t
Outpu
t
Execute action or
search knowledge
base
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

11. Amazon Bedrock
Agents
Enable generative AI applications to
execute multi-step business tasks
using natural language
•Uses power of LLM’s to prompt
and respond using natural
language
•Breaks down and orchestrates
tasks
•Completes tasks by dynamically
invoking APIs
•Securely and privately accesses
company data
•Surfaces chain-of-thought trace
and underlying agent prompts
Generally available
Feature
s
11
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

12. Agent basics
Instructions: “you are an agent that …”
Agen
t
Knowledge
Bases Amazon
Bedrock
do this for
me…
Done. Here’s
the result…
Action
s
12
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

13. Agents use actions to get work done
Instructions: “you summarize meetings and send
results”
Meeting
assistant
list the action
items from my
meeting
Here are the
action items: 1/
…, 2/ …
send them
to my team
Email
sent
Meeting Actions Utility Actions
List Meetings
In: date range Out:
[date, subject,
meeting ID]
Get Action Items
In: meeting ID Out:
[action items]
Send Email
In: subject, recipients,
body Out: status
Get
TeamOut: list of team
member email
addresses
13
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

14. Agents can combine Actions and Knowledge
Bases
Instructions: “you are an HR agent, helping
employees understand HR policies and manage
vacation time”
HR Policy Assistant v2
how much vacation do I
get per year?
as a full-timer with 3
years tenure, you get
15 days
cool. I’d like to take off
December 8 to 15
approved, enjoy. you have
8 more days available
HR Knowledge Base HR Actions
Vacation Policy
Contains the entire
vacation policy for
the company
Request Vacation
In: start date, end
date Out: approval
status,
remaining balance
14
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

15. Agents build on existing enterprise resources
HR Time
Off
Agent
HR Policy docs
Vacation Actions
get-Vacation-Balance
Leave of absence Actions
r
e
q
u
e
s
t-L
O
A
Vacation
microservic
e
Vacatio
n
databas
e
Leave of absence
(LOA)
database
HR Knowledge
Base
search-HR-policy
Existing
resources
15
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

16. Agents can be deployed and invoked from
any app
17
Agen
t
draft
Agen
t
builde
r
Agen
t
user
Even
t
Deploye
d
Agent
Agent
consol
e or
SDK
Custome
r app

Building and testing agents Using production
agents
To deploy an agent, you create a new
Alias, and optionally a new Version
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

17. Agent orchestration is transparent – Trace
Detailed orchestration trace in the console and from the
SDK
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

18. Example: Observability in Natural Language
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

19. Example: Observability in Natural Language
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

20. for Agents
Enable Agents to keep and
use summaries of prior
interactions over time
Build Agents that learn from
previous interactions for more
seamless conversations over time
Enable more personalized
experiences
Use unique memory ids for
separation between users
Retain the summaries for up to 30
days
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

21. for Agents
Enable code interpretation
capability to allow agents
to generate and execute
code to perform complex
analysis
Enable Agents to generate and execute
code to answer questions and solve
problems
Generate charts and analysis
automatically using generated code
Analyze files automatically with
generated code, including CSV, XLS,
JSON, DOC, HTML, TXT, and PDF
Code runs in an isolated environment,
with built-in guardrails enabled
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

22. Multi Agent Orchestration
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

23. Amazon Bedrock Prompt Flows
V I S U A L I Z E A N D A C C E L E R A T E G E N E R A T I V E A I D E V E L O P M E N T W O R K F L O W S
Drag-and-drop
interface
Direct testing
and
deployment
Version control
and aliasing
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.

24. Building AI Agents with Amazon Bedrock
Automates
orchestration of
multi-step
tasks
Provides secure
access to
enterprise data
and APIs
Provides
fully
managed
infrastructur
e
Simplifies
building and
deploying AI
assistants
Lets you choose
implementatio
n languages
25
© 2024, Amazon Web Services, Inc. or its affiliates. All rights
reserved.