Designing custom AI agents for business

In 2 days, you go from a single API call to an agent that carries out a business task end to end. You leave with your own agent, built on your documents and systems: typed tools, sourced search over your corpus, guardrails, a runnable evaluation set. The production readiness document is written during the session.

2 days, 14 hours In person, remote or hybrid IT and development
Quote or registration See the program
Two colleagues draw a diagram of connected modules on a whiteboard

This training is designed for:

The developer wiring in a model

They already call a model from their application and want to hand it a complete task while keeping control over what it triggers.

The IT architect

They have to decide which data and services their organization exposes to an agent, and with what rights.

The lead stuck at the prototype stage

Their agent runs in demos and stalls as soon as cost, logs, failures and human review come up.

Designing custom AI agents for business

Learning objectives

The learning objectives of this training can be assessed. By the end of the session, each participant is able to:

  • Specify a business agent by defining its scope, inputs, expected outputs, tools, success criteria and failure cases
  • Implement a tool-using agentic loop in Python or TypeScript, with typed tool calls, a stop condition, execution caps and step logging
  • Expose your organization's systems and data to an agent through MCP servers and dedicated tools, on the principle of least privilege
  • Build a retrieval-augmented generation (RAG) pipeline on your document corpus and produce sourced answers, with explicit handling of information missing from the corpus
  • Put an agent's guardrails in place: schema-based output validation, per-tool permissions, human approval of actions with side effects, execution caps and an audit log
  • Orchestrate several specialized agents using a pattern you can justify (chaining, routing, orchestrator and workers) while controlling the context passed between them
  • Instrument an agent in production: structured traces, cost, latency, error recovery and regression tests on real cases
  • Draw up a 30-day production readiness document with acceptance criteria, a permissions matrix and measurable tracking indicators

Audience and prerequisites

Audience. Developers integrating language models into business applications. Software and technical architects responsible for scoping the introduction of agentic systems into existing IT systems. Data engineers and platform engineers who need to expose internal data and services to agents. Technical managers and leads who code and drive these systems into production.

Prerequisites. This training is for professionals who write code. The trainee uses Python or TypeScript regularly, can consume an HTTP API and read technical documentation in English, and works with version control. These are entry requirements, checked during the placement assessment. The trainer walks through the code excerpts used in the session and provides a project skeleton, so the time goes to the agent's architecture. Previous experience integrating a language model into an application, even if limited to a single API call. Basic knowledge of information retrieval: lexical index, vector embeddings, similarity measures. Experience with a containerized environment or a continuous integration pipeline.

Each participant's actual level is checked before the session with a placement questionnaire, which is used to adjust the outline.

Program

The full outline, sequence by sequence, is in the detailed program you receive before registering.

Map out the 3 execution modes (single call, scripted chain, tool-using agentic loop) and state the criteria for choosing between them. Observe, on one of your own cases, an agent's full chain from the request to the effect it produces.
Write a tool's contract: typed input schema, a description the model can use, return values and error messages. Choose between a tool implemented in the agent's code and a reusable MCP server.
Describe the RAG pipeline and choose between RAG, tool-based search and long context. Build an ingestion and indexing pipeline on your own professional document corpus.
Identify the attack surface specific to a tool-using agent and the countermeasures that apply. Implement output validation, per-tool permissions and human approval of actions with side effects.
Build an evaluation set from the real requests the trainee handles. Combine deterministic criteria and model-based grading to measure quality and faithfulness to sources.
Choose an orchestration pattern suited to your target task and argue for that choice in terms of cost and latency. Specify subagents with a role, allowed tools, an output format and a completion criterion.
Instrument an agent so every run is readable: structured traces, cost, latency, tool failures. Control how the system behaves under outages, off-schema responses and hostile inputs.
Write up the technical, organizational and legal conditions for putting your agent into service. Assess the learning against the learning objectives.

Want this training for your team?

Answer 3 questions and your estimate appears. We call you back within 48 hours with the quote, the detailed program and possible dates.

Quote or registration

Teaching methods and resources

The training combines demonstrations with commentary, guided workshops and hands-on work on the participants' real cases. Theory is limited to what you need to put things into practice.

  • A practitioner trainer who works on the subjects taught
  • Course materials for each participant
  • Each participant works on their own computer, with their own files
  • Environment set up in advance and tested before the session

In person

The training takes place at the client's premises. The client provides a room with a projector or screen, an internet connection and a computer for each participant.

Remote

The training runs as a live virtual classroom, in a group with the trainer, on a fixed schedule. Participants interact by voice, screen sharing and chat, in both directions and throughout the session. Each participant's connection is logged and serves as the attendance record. Technical support is available by email and phone for the whole session, and a connection test is offered ahead of the day.

Hybrid

A single session can bring together on-site and remote participants. The trainer leads from the in-person venue, and remote participants follow in the virtual classroom with the same individual support and the same deliverables.

Assessment and certification

  • Placement questionnaire before the training starts
  • Continuous assessment through the exercises completed during the session
  • Final assessment of learning against the objectives listed above
  • Satisfaction questionnaire at the end of the session, then again a few weeks later

The training leads to an end-of-training certificate stating the objectives and the assessment results, and to a certificate of completion sent to the funder where applicable.

Jessy knows his subject inside out and, above all, shares it passionately with his students. His advice and follow-up will certainly help the CAMPUS 2023 apprentices and the clubs where they work. I very strongly recommend him.
Christophe Peyrat
Christophe Peyrat

Trainer, ADPS

Official title in the training program: « Concevoir des agents IA métier sur mesure : orchestration multi-agents, recherche augmentée sur vos données, outils et garde-fous » (Designing custom AI agents for business: multi-agent orchestration, retrieval-augmented generation on your data, tools and guardrails). Page updated on August 31, 2026.

Jessy Martin in the free mini course on getting started with Claude

Free course to get started with Claude

Judge our teaching before you commit a budget.

  • 1 hour 20 minutes of video in 5 short modules
  • Installing Claude and setting up your workspace
  • 6 ready-to-copy skills and starter files
  • The privacy settings to apply from your first day
  • A selection of 20 AI tools for marketing
Discover the free course

The course videos and emails are in French.

Questions?

Jessy Martin Academy is a registered training provider [a training organization declared to the French State], so your company can charge the cost to its skills development plan [the training budget a French employer manages for its staff]. Funding from an OPCO [French skills operator that funds employee training] also requires Qualiopi [France's quality certification for training providers] certification, which is in preparation.
The lead time between your request and the start date is set during the scoping call and written into the training agreement [the contract French law requires between the provider and the client]. Allow 15 days minimum, more if an external funder is involved.
Yes. That is exactly what scoping is for: the outline adjusts to your industry, your tools and the participants' actual level, checked with a placement questionnaire before the session.
Let us know in our first conversation. Together we look at adjustments to materials, pace, duration or assessment. The details are on our quality page.
No. Each participant works with their own Anthropic key, in a workspace dedicated to the session that has its own spending cap and is revoked at the end. The credits used during the 2 days are paid by the company, as is the Claude Pro or Max subscription needed for Claude Code.
That is how the training works: the agent is built on the participant's document corpus and systems. Third parties' personal data, sensitive data within the meaning of Article 9 of the GDPR and information covered by a confidentiality obligation are excluded. The trainer checks the corpus before indexing, opens read access on a staging environment and revokes it at the end of the session.

Accessibility · results · complaints

Our quality commitments, our results indicators and the adjustments available to people with disabilities are published on a dedicated page.

See the quality page