AI Agents for Production: MCP, RAG, Evals & Agent Platforms

Understand the components of an agent system and how they work together within defined boundaries.

Explore marketplace roles, business records, policies and architecture before the hands-on demo.

Follow a purchase through delivery and payout, then demonstrate a return and refund across buyer, seller and operator views.

Prepare the course environment: prerequisites, model access, local Docker services and setup verification. Walk through the local setup guide and the optional Google Cloud deployment paths. Download the slides for the diagrams and clickable setup-guide links.

Check access to the Main and Test marketplaces, local Langfuse and the LiteLLM dashboard. Explore the optional cloud Marketplace, runtime selection and tracing paths. Use the linked setup guides for prerequisites, commands, verification and cleanup.

Follow a support question through the Marketplace application, OpenAI Agents SDK, model gateway and read-only MCP tools. Inspect policy retrieval, same-conversation context, current business records and Langfuse traces, then compare the answer after a new purchase. Download the slides and open Exercise materials for the practice instructions and separate reference answers.

Connect the first-agent demonstration to the implementation: request handling, agent construction, MCP tools and evidence. Practise Marketplace journeys and support investigations using your own current records, then compare your observations with the separate reference answers. Open Exercise materials for the shared worksheet and optional cloud extension.

Understand how models and an LLM gateway provide unified access, aliases, permissions, visibility and reliability controls.

Explore LiteLLM, switch the Test model mapping, inspect traces and practice aliases and response caching.

Distinguish stable instructions from dynamic context and see how catalog definitions, records and policy evidence shape an answer.

Inspect context across model calls, compare suppressed sources and practice tracing a complete investigation.

Understand tool contracts, MCP discovery and invocation, local and HTTP transports, and trusted execution boundaries.

Inspect marketplace MCP calls and their inputs and evidence, then add and test a seller-balance tool in an isolated lab copy.

Understand retrieval and answer generation, RAG architecture choices, evidence quality, and how policies combine with business records.

Follow a RAG call, compare order eligibility under two policy versions, and compare paragraph and overlapping chunks with one or four passages.

Distinguish stored conversation history from the bounded context sent to the model. Explore session identity, tool-result retention, context selection and evidence freshness, including the responsibilities of the application and agent runtime. Download the slides for the lesson diagrams and references.

Inspect how earlier messages and tool evidence appear in a follow-up investigation, then compare behavior in the existing Test workspace. Open Exercise materials and start with exercise.md. Record the context and source evidence you actually observe before consulting the separate reference answers.

Understand the agent loop: model calls, authorized tool execution, returned observations and stopping conditions. Compare SDK-managed execution with an explicit loop and examine how limits and recovery affect behavior. Download the slides for the lesson diagrams and references.

Trace normal execution, a simulated business-query failure and recovery, and a constrained stopping case in the Test workspace. Open Exercise materials for the learner task and separate reference answers, including the supplied minimal-loop code. Inspect the actual model and tool sequence rather than inferring success from a final answer alone.

Separate the agent loop, SDK and runtime from the infrastructure that hosts them. Compare self-managed and managed execution, then map the course's local Docker and optional Cloud Run deployments. The revised slides include the optional Presidio Analyzer and Anonymizer containers used by the gateway privacy demonstration.

Distinguish tracing of individual investigations from monitoring across runs. Connect model and tool observations with latency, token usage, estimated cost and feedback, and use those signals to identify a run that needs investigation. Download the slides for the lesson and references.

Investigate recorded runs in your existing local Langfuse lab. Open the Exercise materials resource directory and start with exercise.md. Try the activity before opening solutions/answers.md. No new model calls or business-data changes are required. The learner exercise begins at 10:29 in this video.

Build an explicit quality bar using test cases, independent reference answers, human review, model-based grading and code checks. Interpret scores and disagreements in context, and distinguish a completed upload from an evaluation that has actually run. Download the slides for the lesson and references.

Use your existing Course Setup lab and local Langfuse project. Open the Exercise materials resource directory and start with exercise.md. The directory contains the worksheet, supplied answers, evaluator code, helper scripts and a separate solutions folder. Consult the solutions after your attempt. No fresh model calls are required for the learner tasks. The learner exercise begins at 8:04 in this video.

Examine the boundaries that protect agent inputs, outputs, tools and business data. Compare privacy masking, configured blocking and prompt-injection screening, and distinguish model behavior from enforced authorization. Download the slides for the control map, examples and references.

Compare privacy masking, configured blocking and prompt-injection screening in the existing Test workspace, and inspect their enforcement points in Langfuse. Open Exercise materials and start with exercise.md; companion evidence and separate solutions are in the same directory. Follow the guardrail prerequisites and use synthetic test data. The learner exercise begins at 7:40.

Compare routing policies, classifiers and decision scope. Examine the tradeoffs among answer quality, model cost and routing overhead, using bounded historical experiments and the course's Test workspace controls. Download the slides and evaluate conclusions against comparable tasks and observed evidence.

Compare routing classifiers and conversation versus question scope in the existing Test workspace. Inspect the requested policy, selected model, routing scope and decision cause in Langfuse. Open Exercise materials and start with exercise.md; try the supplied comparisons and heuristic v2 before consulting solutions/answers.md. The learner exercise begins at 9:08.

Compare fixed workflows, model-directed agents and hybrid designs. Explore specialist roles, generator/reviewer correction and the conceptual agent-swarm topic, while distinguishing those patterns from the behavior implemented in the Marketplace lab. Download the slides for the diagrams, examples and references.

Inspect the generator, reviewer and correction stages in the marketplace demo, then practise checking seller payment and payout details. Learner exercise starts at 6:40. Open Exercise materials for the worksheet and separate reference answers.

Distinguish short-term conversation context from long-term semantic, episodic and procedural memory. Explain memory formation, storage, retrieval, authority and lifecycle, and compare the internal implementation with hosted Mem0. Download the slide PDF for the lesson and clickable references.

Observe same-chat and fresh-chat memory, compare explicit internal saving with Mem0 extraction, and inspect retrieval, editing and forgetting. Learner exercise starts at 7:59. Open Exercise materials for the worksheet and separate reference answers. Historical investigation checkpoints are episodic memory; reusable investigation checklists are procedural memory.

Map familiar agent-system components onto Google's agent platform and distinguish SDK development from managed execution. Explore the basic support-agent path, the optional full-cloud Marketplace integration, and the responsibilities retained by the application team. Download the slides for the architecture and setup references.

Inspect the Google support agent in its native Playground and follow the learner practice using the supplied fictional order facts. Open Exercise materials for the setup guide, instructions and separate reference answers. The full-cloud Marketplace integration is an optional extension with its own prerequisites and verification limits.

Map an integrated agent platform through Lyzr's control plane, agent execution, knowledge, tools, models and telemetry. Distinguish native platform capabilities from the application and data services your team still owns. Download the slides for the stack diagram and setup references.

Explore the Lyzr Basics agent, uploaded knowledge and native Playground behavior, then follow the learner practice. Open Exercise materials for the setup instructions and separate reference answers. The optional Marketplace extension connects a separate Lyzr-hosted agent to your existing full-cloud application and authenticated business tools.

Compare cloud, model-provider, framework-associated and integrated agent platforms. Use requirements, component ownership, total operating cost and representative trials to build a shortlist and make a conditional recommendation. Download the slides for the lesson and clickable references.

Apply the decision method to the fictional Company D and Company E requirements. Shortlist two options per company, compare ownership, constraints and costs, then write a conditional recommendation and a validation plan. Open Exercise materials for the worksheet, requirements and separate reference answers. No coding or deployment is required.

Plan the move from a working agent to production: governed data, identity and authorization, human escalation, controlled releases and rollback, cost and capacity limits, incident ownership, and continuous evaluation. Review the evidence needed for release and distinguish platform capabilities from your team’s responsibilities. Download the slides for the lesson and clickable reference.