Présentiel et distanciel
4 days (28 hours)

Building with the Claude API and Agents

A developer course: Messages API, streaming, errors and limits, cost control (batch, prompt caching), structured outputs, tools, MCP, Agent SDK, security, evaluation and deployment. Running project: a support assistant connected to a demonstration ERP.

Training objectives

  • Structure and troubleshoot a call to the Messages API.
  • Read stop reasons and handle streaming, errors, retries and rate limits.
  • Control costs: model choice, effort and thinking, batch, prompt caching, context management.
  • Obtain reliable structured outputs and work with documents, files and citations.
  • Give Claude tools, develop an MCP server and choose between a workflow and an agent.
  • Build an agent with the Agent SDK and customise Claude Code (hooks, permissions, subagents, skills).
  • Secure, evaluate, version and progressively deploy a Claude application.

Target audience

  • Developers and tech leads.
  • Software architects.
  • Data engineers.
  • Integrators.

Prerequisites

  • Working knowledge of Python or TypeScript.
  • HTTP, JSON, Git and the command line.
  • ECI-IA-101 or equivalent experience recommended.
  • A workstation with Python 3.11+ or Node.js 20+ and VS Code.

Detailed program

Day 1 — API foundations

  • The Messages API: anatomy of a call and its response, headers, required parameters, system prompt, content blocks, roles and a stateless API, usage object and request identifier. Lab 1: first call, deliberate errors and a logging wrapper
  • Stop reasons and streaming: stop_reason values, including refusal and context window overflow, checking before any JSON parsing, SSE events and delta types, possible error mid-stream. Lab 2: multi-turn streaming chatbot with max_tokens handling
  • Errors, retries and rate limits: error codes to fix or to retry, spend cap, exponential backoff with jitter, request and token limits, gradual ramp-up, tool idempotency. Lab 3: resilient client (retries, circuit breaker, traffic smoothing)

Day 2 — Data, context and costs

  • Prompt, context and structured outputs: where to place instructions according to trust, XML tags and examples, context engineering (compaction, context editing, memory, tool search), JSON Schema and strict tools, business validation. Lab 4: extracting 20 invoices into validated and checked JSON
  • Documents, files and citations: images and PDFs, Files API and isolation by workspace, native citations and RAG. Lab 5: the assistant's document base with cited answers
  • Models, thinking, batch and caching: cost per successful task and small or large model routing, thinking configuration across model generations, version pinning, batch processing, prompt caching and what invalidates it. Labs 6 and 7: migrating the thinking configuration to a recent model, measuring the gain from caching and batch

Day 3 — Tools, MCP and agents

  • Tool use: defining a tool (name, description, input schema), the tool_use and tool_result loop, tool errors and parallel calls, tool_choice, server tools (web search, code execution). Lab 8: ERP tools (order, stock, ticket) and the complete loop
  • Model Context Protocol: host, client and server, JSON-RPC 2.0, the tools, resources and prompts primitives, stdio and Streamable HTTP transports, security and authentication, configuration in Claude Code and the API's MCP connector. Lab 9: MCP server for the demonstration ERP
  • Workflows and agents: workflow with known steps or agent with unknown steps, five patterns (chaining, routing, parallelisation, orchestrator-workers, evaluator-optimiser), the agentic loop and guardrails, deployment modes. Lab 10: small or large model routing and orchestrator

Day 4 — Security, engineering and production

  • Agent SDK and advanced Claude Code: Agent SDK (built-in tools, subagents, sessions), CLAUDE.md, hooks, deny, ask and allow permissions, skills, non-interactive mode in CI. Labs 11 and 12: assistant ported to the Agent SDK, PR review with Claude Code in CI
  • Security and privacy: direct and indirect injection, defences through least privilege, isolation, channel separation, filtering, human validation and adversarial testing, secrets management, personal data and retention, compliance with Laws 09-08 and 2013-450, multi-tenant risks. Lab 13: attacking then fixing the assistant (5 injections)
  • Engineering, evaluation and deployment: REST, idempotency, pagination and versioning, configuration outside the code, a changed prompt is a deployment (version, test, canary, rollback), robustness, test set, thresholds, LLM judge and traces. Lab 14: 30-case evaluation set, then scenarios and final project

Format and assessment

  • On site or live online, 70% hands-on; in-house sessions on request
  • Delivered in French (official documentation in English)
  • 14 labs, 10 scenarios and 1 graded final project
  • Assessment: quiz (15%), validated labs (35%), scenarios (10%), final project that is functional, evaluated and secured (40%)
  • ECINTELLIGENCE training attestation awarded from a score of 60%

Materials provided to participants

  • ECINTELLIGENCE course manual and lab workbook (PDF)
  • Git repository of the running project (reference code for each step)
  • API cheat sheet: parameters, stop reasons, error codes, usage counters, limits
  • Templates: system prompt, tools, MCP server, evaluation set, hooks
  • 10 scenarios with model answers
  • Dated state-of-the-art appendix (models, pricing, beta features, retirements)

Next steps

As models, versions and features evolve quickly, the content is updated before each session. Product names mentioned (PyTorch, OpenShift, Kubernetes, Proxmox, NVIDIA, Hugging Face…) are trademarks of their respective owners.

Independent training designed by ECINTELLIGENCE. Claude and Anthropic are trademarks of Anthropic. This training is neither delivered nor endorsed by Anthropic and does not prepare for any official Anthropic certification.

Certification

At the end of this training, you will receive a certificate of participation issued by squint.

Price on request

Duration

4 days (28 hours)

Format

Présentiel et distanciel

Next session

On request

Request a quote

Other training courses that might interest you

AI Landscape for Decision-Makers

Présentiel et distanciel

A jargon-free day for decision-makers: an overview of predictive, generative and agentic AI, what makes a project succeed, how to …

1 day (7 hours) Learn more

Claude for Professionals

Présentiel et distanciel

A hands-on course for every profession: writing effective requests, choosing the right model, making the most of projects, documents, search …

2 days (14 hours) Learn more

Deep Learning Fundamentals with PyTorch

Présentiel et distanciel

A hands-on PyTorch course: tensors, GPU computing (CUDA and Apple MPS), autograd, the training loop, evaluation, convolutional networks, then an …

3 days (21 hours) Learn more

Ready to develop your skills?

Join hundreds of professionals who have trusted squint for their skills.

View all our training courses