Agentic AI: Production Grade AI Agents using CrewAI and AWS

Trace the evolution from pre-lms approaches to large language models, including transformers and attention, and explain how these enable agentic ai to plan and act, with hands-on cursor setup.

Explore how generative AI creates novel content—text, images, code, and video—driven by transformers, massive data, and powerful GPUs. See its impact across marketing, software development, healthcare, and creative arts.

Explore how attention drives the transformer architecture, introducing self-attention that enables parallelism and superior handling of sequential data for language tasks and beyond, including translation and large language models.

Explore prompt engineering, where prompts guide AI by setting background, constraints, and goals. Apply the race framework—role, action, context, expectation—to turn vague prompts into effective ones.

Explore how agentic ai autonomously decides, acts, and learns to achieve goals with minimal human intervention, with end-to-end use cases in software development and healthcare.

Explore enablers for agentic AI, focusing on the LLM's reasoning and planning. Use reflection, self-correction, and the react pattern to loop through thought, action, and observation with external tools.

Explore Cursor IDE's AI-powered code editor to generate Python code, run tests locally, and validate inputs while building a Fibonacci last-three-numbers class.

Explore the shift from rule-based systems and NLP to generative AI, transformers, and prompt engineering. Discover how Ray's framework and the React pattern enable agentic AI.

Unveil the purpose of agentic AI by exploring planning, reasoning, tool use, and task execution. Introduce crew AI for autonomous agents and use LLM on AWS bedrock to study observability.

Explore an ai agent example using system prompts, user prompts, and tools to fetch latest product data. See the react pattern, memory, and tool-driven reasoning that produce final responses.

Explore a multi-agent system where an intent detection agent routes queries to specialized products, orders, and returns agents, addressing tools, memory, router patterns, and prompt injection risks.

Explore agentic frameworks that structure autonomous AI agents, enabling planning, reasoning, tool use, memory management, and task execution, with a focus on Cru AI and five popular options.

Explore Crew AI, an autonomous agent framework that lets you define a team of agents with roles and goals, orchestrating tasks through crews and flows for structured, collaborative intelligence.

Demonstrates building a two-agent crew to research a topic and generate a markdown report, using Nolapro LLM via crew AI to produce ten bullet points and a detailed report.

Demonstrates converting reports from markdown to JSON for structured output, showing how to define sections with topic overview, key developments, and impact for machine readability.

Learn how agent observability uses melt—metrics, events, logs, and traces—to analyze internal agent behavior, improve explainability, performance, and cost with traces and Lang Fuse.

Explore how to instrument a crew AI with Lang Fuse to trace execution, visualize spans, measure latency and LM token cost, and improve efficiency.

Explore agentic AI frameworks, focusing on planning, tool use, and multi-agent orchestration with crew AI, building a structured, observable agent app that outputs JSON and markdown reports.

Explore deploying agents with AWS Agent Core, learn version management, session isolation, and a custom authorizer to secure identity management, with emphasis on observability through metrics and traces.

Explore how AWS Agent Core enables secure, serverless deployment of production-grade agents with modular identity, memory, built-in tools, and gateways.

Expose an http endpoint via Bedrock agent core to run the crew ai research and reporting agents, producing a json report from a research topic.

Deploy an agentic AI app to AWS Agent Core, using an HTTP endpoint and IAM authentication; auto scales with sessions and uses Bedrock LM, CodeBuild, and observability.

Explore deploying the agentic application on AWS from Agent Core, with Docker images, ECR, and an IAM role. Observe CloudWatch observability and versioned endpoints with LM calls to Bedrock.

Learn to replace AWS IAM with Cognito for non-AWS users by setting up a Cognito user pool and client, generating a bearer token, and authorizing the Agent Core.

Explore AWS agent core capabilities for building, deploying, and operating production-grade agents with security, scalability, session isolation via micro VMs, observability through CloudWatch, and version management for safe updates.

Explore retrieval augmented generation and context engineering, outlining its necessity and building blocks like chunking, embedding models, vector databases. Apply hands-on with AWS S3, Pinecone, and AWS Knowledge Base.

Explore how vector databases store embeddings and enable similarity-based search for rag workflows, using 1024-number vectors and pine cone's free serverless option.

Learn how context engineering augments language models by supplying curated data, instructions, and tools to overcome knowledge cutoff and black-box concerns, improving reliability and alignment.

Discover retrieval augmented generation, which fetches relevant data from a data store, augments the prompt, and guides content generation in LLMs.

Understand rag data flows, from pre-processing that parses PDFs into text and images and stores chunks in a vector store, to embedder, retriever, and augmented generator for the LLM.

Learn how AWS knowledge base reduces rag complexity by orchestrating pre-processing—from parsing and chunking to embeddings and vector storage—and streamlining retrieval, prompt embedding, and chunk reranking for agentic AI.

Create an S3 data source for rag pre-processing by uploading two AI-generated reports on Agentic AI revolution and quantum computing to a reports folder in a new bucket.

Create a vector database using pinecone to store pre-processing results, configure a 1024-dimension index with cosine similarity, and manage the API key in AWS Secrets Manager.

Create an AWS knowledge base by parsing S3 data, chunking content, embedding chunks, and storing them in a Pinecone vector store for retrieval in rag.

Explore how agentic applications use retrieval augmented generation to choose rag data stores, retrieve data via tools, and augment prompts for the language model to generate the final response.

Connects data sources with a knowledge base and Pinecone vector store to demonstrate a real-time agentic rag workflow, using retrieval tools, LM prompts, and trace analysis.

Learn how Rag uses vector databases to index and retrieve chunks, and how the LM prompt is augmented with retrieved content using Pinecone and the AWS knowledge base.

Explore integrating agentic systems with existing tools and MCP, enable web search for context, push outputs to S3 via MCP, and use the agent core MCP gateway for enterprise-grade adoption.

Master context engineering to curate data, instructions, and tools for better LM reasoning and alignment. Apply the react pattern—think, act, observe—in iterative loops using web search and external updates.

Demonstrate using the Tavli web search tool to augment the LM prompt with public data on emerging technologies, replacing the knowledge base tool for research on electric cars in 2025.

Learn how the model context protocol (MCP) centralizes tool hosting on MCP servers, enabling agents to discover and securely access tools via MCP clients, reducing credential sprawl and code changes.

Learn the MCP architecture, an open source standard using JSON-RPC over streamable HTTP or standard input/output to connect server and client, list and call tools, and manage resources.

Learn to use a remote MCP server to power a search tool (web search, extract, crawl) within crew AI app, connecting to MCP inspector and integrating with the LM workflow.

Explore how the MCP gateway acts as a reverse proxy to centralize credentials and targets, supporting remote MCP servers, rest api, AWS Lambda, and fine-grained authorization.

Demonstrates routing the Tavalisse search tool through the MCP gateway, configuring outbound API key and inbound OAuth, and updating the crew AI app to fetch tools via the gateway.

Publish a research report to AWS S3 by deploying a Python Lambda, granting S3 write permissions, and adding it as an MCP gateway target for the crew AI app.

Explore context engineering and tool integration to address tools integration challenges. Review the Anthropic MCP gateway for no-code or low-code multi-integration, and examine the MCP inspector and demos.

Discover how to empower agentic applications with memory by exploring short-term and long-term memory, their purposes, and practical demos using AWS Agent Core Memory service.

Discover why agents need memory to converse and collaborate with humans, remember past requests, learn from past decisions, and adapt plans to human preferences and changing environments.

Demonstrates converting a single crew into a flow of crews for conversational agentic AI, with check intent, follow-up routing, and memory considerations.

Explore how short-term memory captures conversation in a session to maintain immediate context, and how context window size and strategies like sliding window, summarization, and pre-processing optimize lm input.

Demonstrates using AWS agent core memory to store and retrieve session-based short term memory with seven-day event expiration, enabling follow-up questions with conversation history and a retrieval-augmented generation flow.

Long term memory stores structured information from conversations across sessions, persisting key insights like user preferences, semantic memory, and summaries for proactive, personalized interactions.

Explore how to use AWS agent core to store user preferences in long term memory across sessions, while short term memory handles conversation history and extraction strategies.

Explore memory in genetic applications, from short-term to long-term memory, including how long-term memory stores user preferences and extraction strategies like summarization and semantic memory with crew AI flows.

Explore genetic patterns as design patterns for agentic AI, including plan and execute, orchestrator, worker, and parallel execution, for single and multi-agent systems.

Explore four popular single agent patterns for agentic AI—tool use, agentic rag, react, and plan and execute—highlighting how they reason, act, and adapt.

Demonstrate plan and execute pattern to bring task execution predictability by planning with a planning LM before triggering researchers and reporting agents, contrasting with the react pattern.

Learn to orchestrate multiple agents with key multi-agent patterns, including sequential execution, parallel execution, router pattern, and orchestrator worker pattern, highlighted through practical demos.

Showcases parallel execution in an agentic AI workflow by generating banner images for report sections in parallel with a banner crew and Dall-E, then aggregating into a markdown report.

Demonstrate the orchestrator worker pattern, a hierarchical task decomposition where a manager agent splits tasks into subtasks and delegates to intent analyzer, researcher, and report analyzer.

Compare structured workflow and autonomous orchestration as design philosophies, contrasting deterministic, event-driven processes with autonomous, non-linear agent collaboration for production-grade automation and exploratory tasks.

Explore single agent patterns like tool use, agentic, rag, and react. Study multi-agent patterns such as sequential and parallel execution and compare autonomous orchestration with structured workflow philosophies.

Discover how agent evaluation raises output quality in enterprise-grade applications by exploring LM evaluation, benchmarks, online and offline methods, and inline critique agents.

Explore agent evaluation and its four types, including LM evaluation with benchmark results, online evaluation after deployment, live feedback during execution, and offline evaluation like unit and integration testing.

Learn how lab evaluations use Elm benchmarks to assess a language model's accuracy, coherence, relevance, and safety, with MLU and GPC as examples, and note agentic evaluations require custom data.

Online evaluation probes apps in live production, measuring success rates and user satisfaction with real time traffic while comparing LM as a judge and human annotation for scoring.

Demonstrate using an llm as a judge for online evaluation, applying evaluators to score report quality and hallucination across traces, with trace-based scoring and dashboards.

See a demo of human annotation, filtering traces over 90 seconds and annotating execution plans as not a good plan, somewhat optimal plan, or optimal plan to improve insights.

Learn how inline evaluation scores an agent's real-time performance, provides feedback to adjust the current execution, and uses a critic agent pattern with multi-agent critique to improve output quality.

Demonstrates the critique agent pattern within an end-to-end production-grade AI workflow. Shows inline evaluation, intent analysis, and feedback loops between research and critique agents to improve report quality.

Explore human in the loop and inline evaluation by examining five levels of user involvement, from operator to observer, and how workflow engines complement agentic AI.

Study offline evaluation for production-grade agentic AI by testing in a sandbox with a test set of inputs and expected outputs, measuring latency and LM-based performance scores.

Explore how agentic ai are probabilistic and require evaluation mechanisms, including lm-based online assessments, offline assessments, human annotation, and a critic agent for online feedback.

Explore inter agent communication with a two-way protocol, tracing its origin from Google to the Linux Foundation, and compare it to MCP through demos.

Explore why inter agentic communication matters, distinguishing inter agentic application communication across separate hosted apps from internal agent interactions, and how meshing agentic apps enables a business goal.

Learn about the A-2a two-way protocol, an open standard for seamless, multi-turn inter-agent communication built on HTTP, JSON-RPC, OAuth, and API keys, with agent cards for discoverability.

Explore end-to-end a2a interaction flows, from client agent discovery and message exchanges with remote agents and tools to multi-turn task handling using context id session and message and task objects.

Expose an emerging agentic AI application over eight way with a two-way interface, using an agent skill and agent card to enable multi-turn conversations and run on port 9000.

Explore the agentic ai via a two way protocol using the eight way inspector to fetch agent cards, inspect skills, and generate a research report through multi turn interactions.

Demonstrate inter-agency communication by connecting an LMS virtual assistant to a remote emerging technology research agent through multi-turn conversations with local chroma db memory.

Compare MCP and A-2a; MCP focuses on agent-to-tool integration and context access, while A-2a standardizes inter-agent communication with multi-turn task-based interactions, both coexisting.

Explore HOA inter-agency communication, its open-source origins, two-way protocols, agent discovery with cards, and MCP and A-2a for multi-turn interactions using HTTP, Json-rpc, and OAuth.