Claude & Agentic AI Masterclass — 100 Hours of the Claude API, MCP and Production Agents

• System prompt architecture: role, constraints, output contract, examples
• XML tags, long-context organisation, document-first prompting
• Extended thinking: budgets, when it pays for itself, interaction with tools
• Structured output: JSON schemas, tool-based extraction, validation loops
• Prefill, stop sequences, and controlling verbosity and format
• Lab: prompt regression suite with 40 golden cases

• Tool schemas, tool_choice, parallel tool calls, tool results and errors
• Designing tools an LLM can actually use: naming, granularity, idempotency
• Server-side tools: web search, code execution, computer use — capabilities and limits
• Client-side orchestration loops; the tool-runner pattern
• Failure handling: timeouts, partial results, retries, human escalation
• Lab: a 6-tool assistant with deterministic fallbacks

• Prompt caching: cache breakpoints, TTLs, hit-rate measurement, cache-aware prompt layout
• Batch API for offline workloads; throughput planning
• Model cascades and routing; when a cheap model plus verification beats one expensive call
• Token budgets, context compaction, summarisation strategies for long sessions
• Load testing, SLOs, graceful degradation
• Lab: take a workload from $0.31 to $0.04 per request with measurements

• Why MCP exists; architecture: hosts, clients, servers, transports (stdio, HTTP/SSE)
• Server primitives: tools, resources, prompts, sampling, roots, elicitation
• Building a server in Python and TypeScript; schema design and validation
• Auth patterns: OAuth, tokens, per-user scoping, multi-tenant servers
• Security: tool poisoning, confused deputy, prompt injection through resources, allowlists
• Testing MCP servers; the inspector; publishing and versioning
• Lab: Project 2 — ship a real MCP server with an auth story

• Agent loop anatomy: gather context, act, verify, repeat
• Subagents and delegation; when parallel agents help and when they thrash
• Memory: session state, files-as-memory, compaction, retrieval-backed memory
• Permission systems, sandboxing, allow/deny rules, dry-run modes
• Hooks and lifecycle events for policy enforcement and logging
• Long-running and scheduled agents; resumability
• Lab: Project 3 — autonomous workflow with approval gates

• CLI workflows: planning mode, diffs, checkpoints, worktrees
• CLAUDE.md, project context, skills, slash commands, subagent definitions
• Hooks for tests, linting, secret scanning and policy checks
• Plugins and marketplaces; distributing tooling to a team
• Headless mode in CI: automated review, migration and refactor jobs
• Lab: build a team plugin that encodes your engineering standards

• Long context vs retrieval — measured trade-offs at 200K+ tokens
• Contextual retrieval; citation-first answering; grounding checks
• Connecting Claude to warehouses: Databricks, Snowflake, BigQuery via MCP or tools
• Permission-aware retrieval and row-level security in agent contexts
• Lab: Project 1 knowledge agent

• Anthropic API vs Amazon Bedrock vs Google Vertex AI: features, regions, pricing, IAM
• Secrets, key rotation, per-tenant isolation, egress control
• OWASP LLM Top 10 applied to agentic systems; blast-radius design
• Observability: OpenTelemetry traces, token dashboards, eval-in-production
• Usage policies, data retention, and what not to send to any model
• Lab: deploy the knowledge agent on Bedrock with private networking

• Evaluating agents: task success rate, trajectory analysis, cost per resolved task
• LLM-as-judge rubrics for agent output; regression suites for tool behaviour
• Red-teaming agents: injection through tool results, over-permissioned tools, runaway loops
• Capstone presentation and architecture defence
• Agentic-AI interview questions, portfolio framing and resume rewrite
• Keeping up: reading changelogs, deprecation notices and the MCP spec

No. It is an independent training programme built on Anthropic's public documentation and SDKs. We are not affiliated with or endorsed by Anthropic, and Claude is a trademark of Anthropic.

Because agentic engineering — MCP, tool design, permissioning, context management — is deep enough to deserve it, and the patterns transfer. If you want breadth across providers, take the Generative AI course; many students do both.

Constantly. Content is refreshed every batch and there is a standing 'what changed' segment. Deprecated parameters and models are flagged explicitly.

Roughly $25–$50 across the course. Prompt caching and the small model tier are taught in Module 4 specifically so your own bills stay low.