MCP and Agent SDK training: AI agent architecture for production
Two days to design, evaluate and run your own AI agents: MCP servers, the Claude Agent SDK and its permissions, subagents, reproducible evaluation, security and observability. The natural next step after "Claude Code in production".
Target audienceDevelopers, tech leads, software architects and CTOs who design AI agents for production
Colombani.ai's "AI agent architecture: MCP and Agent SDK" course lasts two days (14 hours), in-house, for one to ten developers, tech leads, architects or CTOs, at your premises or remotely. Participants build an MCP server compliant with the 28 July 2026 revision, build an agent with the Claude Agent SDK while controlling its permissions and cost, add a verification subagent, then make it measurable, safe and observable. It is aimed at teams who have taken "Claude Code in production" or have an equivalent level. Colombani.ai is a Qualiopi-certified training provider (certificate QUA009472).
Updated
What you will be able to do after the MCP and Agent SDK training
Choose the architecture suited to a need: control, operations, cost, risk
Build an MCP server compliant with the 28 July 2026 revision
Build an agent with the Claude Agent SDK whose permissions are genuinely restricted
Have a main agent work with subagents, and check their results
Evaluate an agent reproducibly, before and after every change
Secure an agent and its MCP servers
Track in production what an agent does and what it costs
MCP and Agent SDK training programme (14 hours)
- Day 1, session 1 (1 hour)
Choosing an agent architecture
- The agent loop: model, tools, context, stopping criteria
- Four ways to build an agent: a hand-written loop on the API, client SDK tools, the Claude Agent SDK, agents hosted by Anthropic
- Agent or deterministic workflow: when autonomy adds value, when it adds risk
- What applies to other model providers, and what is specific to Claude
- Day 1, session 2 (2 hours)
Building an MCP server
- The 28 July 2026 revision: stateless protocol, server discovery, requests for user input
- Tools, resources and prompts; descriptions and error messages a model can use
- stdio and HTTP transports; demonstration of OAuth authorisation for a remote server
- Exercise: an MCP server on a sample API, checked with MCP Inspector then used from Claude Code
- Day 1, session 3 (2 hours)
Building an agent and controlling its permissions
- The Claude Agent SDK in Python or TypeScript, connected to the previous session's MCP server
- Permission pitfalls: what an allowed-tools list does not restrict, the mode you must always set explicitly, how to genuinely remove a tool
- Validation by code and hooks; a spending cap set with a margin
- Exercise: an agent that performs a business task with a tool removed, an explicit mode and a spending cap
- Day 1, session 4 (2 hours)
Running exercise: a first working agent
- Short specification: task, inputs, outputs, success criteria, forbidden actions
- MCP server and agent on a supplied sample case, with entirely fictitious data
- Demonstration to the group and peer review
- Day 2, session 5 (1 h 15)
Orchestrating subagents
- Declaring a subagent: instructions, allowed tools, model, number of turns
- What a subagent inherits in terms of permissions
- Exercise: a verification subagent that rejects anything out of scope
- Day 2, session 6 (1 h 30)
Evaluating an agent reproducibly
- Nominal, edge and adversarial cases; checks by code, by a model against a rubric, by a human
- Repeated runs, pass rate and acceptance threshold rather than a single run
- Exercise: a set of eight test cases, three runs each, launched by a single command
- Day 2, session 7 (1 h 15)
Securing an agent and its MCP servers
- Least privilege, secrets out of the code, isolated execution
- MCP-specific risks: a server acting with rights other than the user's, poisoned tool descriptions, an uncontrolled third-party server
- Instruction injection through documents read; human approval of sensitive actions enforced by a hook, not by an instruction
- Day 2, session 8 (1 hour)
Observability
- OpenTelemetry traces and metrics, cost and tokens per request
- Conversation content not exported by default: a choice to justify under the GDPR
- Container deployment and versioning of instructions
- Day 2, session 9 (2 hours)
Running exercise: a measurable, observable and safe agent
- Traces, cost cap, tightened permissions and human approval on the day 1 agent
- Running the evaluation bench and comparing before and after
- Final quiz, production plan, review
MCP and Agent SDK training prices
The training is delivered in-house and invoiced per group, according to group size; adapting the programme to your own work situations is included. VAT not applicable, article 293 B of the French tax code. Invoiced in addition, on the quote: travel costs outside the Paris region and API usage, through your own key or a training key with a spending cap that is revoked at the end of the session (roughly €20 to €50 per participant for the two days). In France, a funding request can be made to the company's OPCO or to the FIF-PL, subject to the funder's decision.
| Group size | Price for the two days |
|---|---|
| 1 to 4 participants, including a single person | €3,500 excl. VAT |
| 5 to 7 participants | €5,250 excl. VAT |
| 8 to 10 participants | €7,000 excl. VAT |
Length, price and funding
2 days (14 hours)
Developers, tech leads, software architects and CTOs who design AI agents for production
Having taken "Claude Code in production" or an equivalent level: everyday development in Python or JavaScript/TypeScript, regular use of a coding agent, REST API calls, Git. A machine with Python 3.11+ or Node.js 20+, Git and installation rights. A placement questionnaire checks these prerequisites before enrolment.
1 to 10 people
€3,500 to €7,000 excl. VAT per group
OPCO funding possible.
Demonstrations followed by hands-on work on a supplied exercise repository, in Python or TypeScript, with a running case on fictitious data that carries on from one exercise to the next. A success rubric is handed out at the start of each exercise.
Technical placement questionnaire before enrolment, six exercises assessed against a rubric, short assessments for sessions without an exercise, a ten-question final quiz. At the end, a before-and-after self-assessment, then the trainer validates each objective (acquired, in progress, not acquired), recorded in the end-of-training certificate. New course: results indicators will be published after the first session.
Ulysse Trin, founder of Colombani.ai, holder of the Claude Certified Architect Professional certification and the Claude Partner Badge: Claude Code, both issued by Anthropic.
At least 2 weeks between enrolment and the start of training.
Qualiopi-certified provider (certificate QUA009472), registered under training provider number 11757549975: OPCO or FIF-PL support can be requested. The decision rests with the funding body, based on your sector, its criteria and its budgets. Colombani.ai prepares the application with you.
Price per group, €3,500 to €7,000 excl. VAT, depending on group size and format: charged per group rather than per person. OPCO or FIF-PL funding can be requested depending on your status; the amount left to pay depends on the sum granted by the funding body.
Receive the programme and a quote
Tailored programme and quote within two business days.
Accessibility
This course is accessible to people with disabilities. Contact the disability officer in advance to discuss possible adjustments.
Ulysse Trin, [email protected]
Post-training support
Frequently asked questions about the MCP and Agent SDK training
How does this differ from the Claude Code in production course? +
"Claude Code in production" teaches how to work with a coding agent on your own repositories. This course goes one step further: it teaches how to design, evaluate and run your own agents, with MCP and the Claude Agent SDK. The former is the recommended prerequisite for the latter.
What is an MCP server? +
MCP, the Model Context Protocol, is an open standard that connects an AI agent to data and tools. An MCP server exposes a system (a database, an API, an application) as tools an agent can call. The standard is supported by several agent environments, not only by Claude.
What is the Claude Agent SDK? +
It is Anthropic's library, in Python and TypeScript, for building an agent based on the Claude Code engine into an application: tools, sessions, permissions, hooks, subagents. The course focuses on its permissions, several of whose settings do not do what their name suggests.
Can the exercises use our own systems? +
The exercises use a supplied sample case with entirely fictitious data. A system from your organisation can be used for a group of four people at most, on a test account or fictitious data, with your written agreement.
Does the course lead to a qualification? +
No. It leads to an end-of-training certificate listing the objectives and the result of their assessment. In France, a funding request can be made to the OPCO or the FIF-PL; it is not eligible for the CPF.
Why are no satisfaction indicators shown? +
It is a new course: its results indicators will be published after the first session, as for every Colombani.ai programme.
Related courses
Advanced Claude Code training: hooks, subagents, MCP and continuous integration
Two days on your repositories to take Claude Code to production: security and quality hooks (PreToolUse, PostToolUse), orchestration of subagents and parallel sessions on worktrees, MCP servers connected to your tools, headless Claude Code in continuous integration (GitHub Actions), permissions and defence against prompt injection.
AI agent security training
The risks of autonomous agents (prompt injection, exfiltration via MCP) and their countermeasures: sandboxing, permissions, security hooks. AI agent red teaming.
Receive the programme and a quote
Tailored programme and quote within two business days.
Got a project in mind?
Describe your situation. The first 30-minute call is free. Get a frank response within one business day.