Claude GuideDevantia × Executive Partners Group
DevantiaExecutive Partners Group
0/23 Jean-Christophe Leroy
Jean-Christophe Leroy
Guide author
✓ Verified on September 3, 2026⏱ 5 min read
Module 15

Building AI agents

In plain terms — what is an agent? An agent is Claude to which you entrust a mission that runs on its own, without you re-issuing the request each time: monitoring every morning, a report every Monday, a first pass at sorting leads. Two levels in this module: a simple use via plugins (accessible to everyone, below), then an advanced infrastructure for your technical teams (the rest of the page — skim it or forward it to your IT department).

🔧 Via Claude plugins (native)

Agents with skills, tools and roles defined in plugins.

⚙️ Via n8n + Claude API

Missions triggered by an event: an incoming email, a set time, a submitted form (technically: "webhooks").

🛠️ For IT — technical details (click to expand) · Claude Managed Agents — the official infrastructure

Claude Managed Agents — the official infrastructure

Section for your technical teams. Anthropic hosts all the "machinery" that runs autonomous agents. Four concepts, by analogy with a mission entrusted to a staff member: who does the work, where they do it, the mission under way, and the exchanges during the mission.

🤖

Agent

The model, its system prompt, its tools, its MCP servers and its skills. This is "who" does the work.

🏗️

Environment

A configured container template (packages, network access). This is "where" the work runs.

⚡

Session

A running agent instance inside an environment, performing a specific task.

📨

Events

The messages exchanged between your application and the agent (user turns, tool results, statuses).

Anthropic manages the sandbox, tool execution and context management. For full control, use the Claude Agent SDK.

📌 Managed Agents — key capabilities Outcomes: you define the success criterion, and the agent iterates until it's met. Memory & Dreams: the agent remembers past sessions and reorganizes its memories autonomously. Native multi-agent orchestration, webhooks across the whole lifecycle and scheduled deployments (cron) — your agents run on a schedule, with no infrastructure to maintain. Cost of the machinery: $0.08 per session hour, on top of tokens.
🛠️ For IT — technical details (click to expand) · Technical setup (for your CTO)

Technical setup (for your CTO)

# Prerequisites # Python 3.11+, Anthropic API key pip install "anthropic>=0.91.0" python-dotenv # Initialization from anthropic import Anthropic client = Anthropic() MODEL = "claude-opus-5" # the frontier at half price — "claude-fable-5-1" for the ceiling, "claude-sonnet-5" for simple tasks
🛠️ For IT — technical details (click to expand) · Multi-agent orchestration — creating a team of Claudes

Multi-agent orchestration — creating a team of Claudes

Most people use a single Claude. Power users orchestrate a team of Claudes in parallel. A main agent (orchestrator) breaks down, delegates and aggregates the work across specialist agents working simultaneously.

Overview of the multi-agent system 🧠 ORCHESTRATOR breaks down · delegates · aggregates 🔍 Research 📊 Analysis ✍️ Writing ⟨/⟩ Technical 📁 SHARED FILESYSTEM — read / write for all agents

The 5 steps

1

Define the orchestrator agent — the conductor that directs without doing the work itself. Clear instructions on breakdown, delegation criteria and expected delivery formats.

2

Create the specialist agents — in .claude/agents/, one .md file per agent (research.md, analysis.md, writing.md, technical.md). Each one defines: name, role, skills, tools, model, constraints, output format.

3

Configure the shared filesystem — a common folder (/projet_partage/ with subfolders 01_brief, 02_donnees, 03_travail, 04_validations, 05_livrables) where all agents read and write. It's their communication channel with no intermediary.

4

Launch your first multi-agent task — hand the orchestrator a complex task (full audit, multi-source report, refactoring). Watch the sub-agents divide up the work.

5

Monitor and iterate — read each agent's logs, identify friction, adjust the instructions. Each iteration improves the team's performance.

🎯 Example — Acquisition due diligence The orchestrator receives: "Analyze this acquisition target." It delegates: the Research agent compiles the history and the market, the Analysis agent dissects the financial statements, the Technical agent audits the IT stack, the Writing agent consolidates everything into an investment memo. It all happens in parallel, on a shared folder.
💡 One Claude agent is good. A team of Claudes that coordinates on its own is a whole different category of tool. Time savings (parallel work), higher quality (each agent brings its expertise), more manageable complexity (ambitious projects), scalable (add agents as your needs grow).
🎯 Executive committee agent examples Monitoring Agent: daily web collection → analysis → summary on Slack at 8am.
Reporting Agent: KPIs Monday → Google Doc report.
Prospecting Agent: Apollo leads → emails → human review.
🛠️ Your turn — 5 minutes

Describe a recurring task to Claude and ask: "propose the trigger, the stop criterion and the guardrail".

Expected result : A three-line agent plan you could hand to your CIO.