This guide does not invent a method: each step is based on projects of AI agents already described in the directory, with their source (business, media or tool editor), and on official library skills. The figures quoted are those of the sources; The editorial staff did not test these agents.
Four ways of doing this come out of the files: setting up an assistant in a mainstream tool like ChatGPT, using a codeless agent platform (Copilot Studio, Dust), chaining tasks with an automation tool (n8n, Make), or developing with a framework (OpenAI Agents SDK, Google ADK, LangGraph).
See also the library of AI agents' skills and the MCP server guide and the app guide and all use cases.
GroupActive gathered the needs of its network during a seminar: the priority was to assist in prospecting (form). Doctolib started with a single scenario, managing access rights to a doctor's agenda (form). At Malt, the team observed the participants' working day and identified the difficult tasks before building with them (form).
Tutiac's assistants were fed into the cooperative's history, product names and language elements. Prospecting officerAlmais connected to Salesforce; that of Montreal draws on more than 40,000 pages and two internal systems; that ofWayflyerGo through Notion, Slack and Jira, and interview internal tools via MCP. To connect a service to an agent by MCP, the skill dBuild a quality MCP serverdescribes the approach, and theserver guide MCPexplains the protocol, its installation and precautions.
At Doctolib, each agent has a role, a specialized prompt and a limited list of tools. ManoMano's prompt specifies which error to prefer: it is better to duplicate than an erroneous reconciliation. For reusable instructions, the SKILL.md format (What's a skill?) stores them in a folder that the agent loads when the task is ready; the skill of the skillCreate and evaluate your own skillsaccompany their writing.
ManoMano has removed to the agent the ticket creation tool: he searches, and a classic code loop creates tickets. Carrefour disconnected the automatic creation of Jira tickets after a month because it produced hundreds of unnecessary tickets. At Doctolib, the model never executes a sensitive action itself: the user confirms with a click.
At Alma, the product manager reads, corrects and publishes the documentation proposed by the agent (form). At GroupeActive, the expert rereads and validates the commercial proposal. AtSAM, TPE of retired counsel, arbitrations, validations and sensitive letters remain supervised by the manager.
The City of Montreal reports 95% accuracy in internal testing and follows its agent in a Power BI dashboard.PMP Strategyconvenes a weekly committee to decide which agents to publish;Bpifranceclassified agents into two levels of risk and listed authorized data connectors. On the skills side, Google publishesEvaluate and improve an AI Officer.
In the wine cellar of Lugny (Saône-et-Loire), the quality, hygiene, safety and environment manager has set up in
Tutiac, Gironde cooperative cellar, has configured several assistants in ChatGPT shared by the marketing team: writing for networks, counter-labels and fact sheets, and one regulatory assistant per importing country.
GroupActive, SME that accompanies a network of 80 independent experts, has created with Copilot Studio an agent that turns notes taken into customer appointments into chapitres commercial proposal.
The City of Montreal has integrated into its site a virtual agent designed with Copilot Studio: it draws from more than 40,000 pages and gives, for example, the waste collection schedule corresponding to the address of the resident.
Danone uses an agent to analyze customer orders, regardless of the format, and cross-checks prices with its internal systems to correct deviations before billing.
Alma, a French split payment solution in a dozen European markets, has built an agent linked to Salesforce who describes opportunities and prepares prospecting messages adapted to each merchant.
At Alma, an agent reads the balance sheet written by the product manager at the end of a project and proposes updating the documentation; the product manager re-reads, corrects and publishes.
Wayflyer, a Dublin fintech that finances growing brands, has built two agents: one prepares the diagnosis as soon as a ticket arrives, the other updates the procedures after its resolution.
Florian Aujames, Commercial Development Officer at Volcos Protection, TPE Electronic Security, trained alone and uses ChatGPT, NotebookLM and Make for its prospecting, documentation and customer tracking.
At ManoMano, a chain of agents analyses the crashes of mobile applications every four hours, looks for whether the problem is already known and links each case to a ticket; the creation of tickets has been withdrawn at the AI.
Carrefour created an agent who learns from the history of his team's conversations "platform engineering" to answer questions already processed on its data platform, before extending it to other teams.
For its support to healthcare professionals, Doctolib has designed Alfred, a system of agents that understands the demand, will search for useful information and prepare the action, which the user confirms himself with a click.
A retired consulting TPE has built a multi-agent system that extracts data from documents, feeds the CRM, tracks missing parts, and revives customers.
Bpifrance opened a codeless agent platform to 500 pioneer collaborators, after a four-month experiment and the establishment of a governance framework with Artefact.
PMP Strategy, a 150-person strategy consulting firm, publishes a library of agents validated by a weekly committee, including an agent who reports the shortcomings and inconsistencies of a presentation before she leaves the client.
At Malt, the central AI team abandoned generic presentations for workshops prepared by observing each team, and provided a platform for secure tools: adoption increased from 55% to 90% according to the article.
skill-creator
Helps to create a zero skill, modify or optimize an existing skill and measure its performance: evaluations, comparisons with variance analysis, optimization of the description to trigger it at the right time.
mcp-builder
Guides the creation of MCP servers that allow a model to act on external services with well-designed tools, in Python with FastMCP, or in Node and TypeScript with SDK MCP.
claude-api
To be used as an up-to-date reference for building an application that calls the API Claude or an SDK Anthropic: model identifiers, parameters, streaming, tool call, MCP, agents, cache, token count and migration between models.
google-agents-cli-onboarding
Cli-agent entry point on Agent Platform, as soon as you want to create, develop, build with the ADK (Agent Development Kit), run locally, debug, test, evaluate, deploy, publish or supervise an agent.
google-cloud-solution-build-deploy-agents
Interactive workflow to design, build and deploy a multi-agent agent or system on Google Cloud: needs collection, architecture adapted, and then deployment instructions, with a validation phase.
gemini-agents-api
The API Managed Agents is the control plan for Gemini Enterprise Agent Platform: it allows you to create, configure, list, update or delete server-side managed agent resources, with system instructions, sandbox files, local or remote skill register and tools, before you open a conversation.
agent-platform-eval-flywheel
This skill measures and improves the quality of a model or agent on Google Cloud, using the Eval Quality Flywheel method.
agents-sdk
The skill builds, debugs or re-reads an application based on the SDK Cloudflare Agents (agents package).
developing-genkit-js
Genkit guide for Node.js and TypeScript: feed, tools, Dotprompt prompts, observability, middleware, and troubleshooting errors, validation or typing.
Yes, according to several records. Tocave de Lugny, the quality manager, trained at the AI, configured a "expert" in ChatGPT.Bpifranceopened a codeless agent platform to 500 employees; atPMP Strategy, each consultant can create his agents. Projects linked to internal systems often involve a technical team:City of Montrealbuilt its agent with its IT team, reusing existing API.
The skills presented here are published in public deposits on GitHub and consult freely; their license, verified on each card, says what you can do with it. Several relate to the construction of agents (list above). The case sheets do not give the price of the mentioned platforms: check with each publisher.
A skill is a folder that contains a SKILL.md file: a header name and description, then instructions, sometimes scripts and resources. According to the documentation of the officer, the agent keeps the name and description on a permanent basis, and only loads the rest when the task is ready; Unlike a quick one, the skill is reused from conversation to conversation. Explanation and sources:What is a skill (SKILL.md)?