La pile d'outils AI Workbench : 7 outils qui transforment l'IA d'une fenêtre de chat en un véritable système de travail

AgentMail: email inboxes for AI agents

"Give your agent its own inbox."

AgentMail is built around one important idea: AI agents need their own inboxes.

Email remains one of the most important interfaces for business. Confirmations, customer replies, vendor attachments, contracts, verification links, support conversations, follow-ups — if an AI agent cannot participate in email, it cannot participate in a large part of real work. AgentMail is an email inbox API for AI agents. It lets developers create inboxes, send, receive, and search messages via REST API — giving agents their own email addresses in the same way Gmail gives humans inboxes.

The key distinction is separation. A human inbox is for human attention. An agent inbox is for agent workflows. You do not necessarily want a support agent, sales agent, research agent, or scheduling agent operating directly from your personal Gmail account. You want it to have a scoped inbox, a clear identity, logs, permissions, and a workflow built around what it is allowed to do.

AgentMail is useful for:

• Support agents that receive customer questions and draft or send replies.
• Sales agents that qualify leads, send follow-ups, and monitor responses.
• Recruiting agents that receive resumes, schedule screens, and follow up with candidates.
• Research agents that subscribe to newsletters, collect reports, and process incoming attachments.
• Operations agents that receive invoices, confirmations, receipts, and vendor updates.
• Onboarding agents that guide new customers through forms, documents, and setup steps.

Any agent that can send email must have clear permissions, restricted scopes, logs, and approval rules for sensitive replies. The direction is right — if agents are going to do real work, they need real addresses.

Obsidian: the private memory layer

"A second brain for you, a first brain for your agents."

Obsidian is usually described as a note-taking app or second brain. In an AI workbench, it becomes something more specific: a private, file-based memory layer that agents can read.

Obsidian stores notes as local Markdown files. That makes the vault readable, portable, scriptable, searchable, and compatible with local agent workflows. For a human, Obsidian is a place to think. For an agent, Obsidian can become the map of your projects, relationships, decisions, tasks, daily notes, and operating context.

Agents perform better when context exists in a durable, readable structure before the task begins. Instead of asking an assistant to infer the day from messy context, the agent can read a structured dashboard and produce a useful briefing — covering what needs attention, what is at risk, which client relationship is stale, and what the most important task is right now.

A strong Obsidian setup for AI work includes:

• Daily notes — what happened, what changed, what still needs attention.
• Project notes — goals, status, next actions, deadlines, and decisions.
• People notes — relationship context, past conversations, and follow-ups.
• Task notes — due dates, priorities, energy level, and linked projects.
• Dashboard notes — reads from the rest of the vault and surfaces what matters now.

With plugins like Dataview, Obsidian can query notes based on properties — type, status, priority, deadline, client — turning a folder of notes into a live operating surface. A dashboard can show today's priorities, active projects, upcoming deadlines, open loops, and client health without manual compilation.

Obsidian rewards structure. If your notes are inconsistent, your dashboard will be incomplete. If properties are missing, agents will miss context. Use Obsidian when the memory is personal, private, local-first, and long-lived.

Notion: the shared structured workspace

"The shared context layer for teams."

Notion occupies a different place from Obsidian. It is less about private Markdown memory and more about shared structured work.

Obsidian is excellent for personal thinking, local files, and agent-readable private memory. Notion is excellent for team collaboration, databases, project views, shared documentation, and operating systems that multiple people need to use. In an AI workbench, Notion is the shared context layer.

A founder can maintain private strategy notes in Obsidian while the team uses Notion for shared projects, public documentation, and operational tracking.

Notion is useful for:

• Company wikis that explain how the business works.
• Project trackers with owners, deadlines, statuses, and milestones.
• Content calendars for newsletters, social posts, articles, and launches.
• Hiring pipelines for candidates, interviews, notes, and decisions.
• Client portals for deliverables, onboarding, and shared documentation.
• SOP libraries that agents and humans can reference before doing work.

The agentic value is straightforward: if the team keeps its structured knowledge in Notion, AI tools can use it to understand the business. A model asked to draft a launch plan should know the product positioning, roadmap, existing campaigns, brand rules, and previous launches. Notion can store that context in a way teams actually maintain.

The risk is that Notion can become too flexible. Databases need consistent fields, pages need templates, and teams need rules for what belongs where — otherwise the AI assistant has access to information but no reliable structure. Use Notion when the memory is shared, operational, collaborative, and structured.

Composio: the integration layer for agents

"The plumbing layer — not glamorous, but no building works without it."

Composio sits in the tool-access layer — and it may be one of the most important tools in the stack.

Agents are limited when they can only produce text. To do useful work, they need to interact with tools: Gmail, Slack, Linear, GitHub, Notion, HubSpot, Google Calendar, Figma, Stripe, Jira, Salesforce, and hundreds of other systems. But connecting agents to tools is messy. Authentication is hard. Permissions are hard. APIs differ. Each tool has its own rules, scopes, errors, and edge cases.

Composio gives AI agents secure access to a large library of tools through MCP or direct APIs, with safer agent actions and reduced risk of destructive behavior or data leaks.

The hard part is permissioning. An agent that can read a calendar is different from one that can delete events. An agent that can draft an email is different from one that can send it. Composio matters because the future of work agents depends on structured tool access: without tools, agents remain advisors; with unsafe tools, agents become a liability; with managed tool access, agents become workers.

Microsoft Copilot: the enterprise work graph

"Strongest where Microsoft is already the operating system of the company."

Microsoft Copilot is strongest where Microsoft is already the operating system of the company.

For many organizations, work lives in Outlook, Teams, Word, Excel, PowerPoint, SharePoint, OneDrive, Planner, and the Microsoft Graph. In that environment, the most valuable AI assistant is the one that understands the internal work graph: emails, meetings, documents, chats, relationships, calendars, files, and permissions.

Microsoft 365 Copilot is built around an intelligence layer that connects individual and organizational knowledge in the flow of work — personalizing responses based on actual work context, relationships, and patterns rather than generic retrieval alone.

Copilot is useful for:

• Summarizing meetings and Teams threads.
• Drafting documents and presentations in Word and PowerPoint.
• Finding information across Microsoft 365.
• Analyzing Excel data and generating insights.
• Working with Outlook and calendar context.
• Helping employees navigate organizational knowledge.
• Building agents in Microsoft's enterprise context.

The tradeoff is ecosystem dependence. Copilot is most powerful when the company already lives in Microsoft 365. If your work is scattered across Slack, Notion, Google Drive, Linear, and custom tools, Copilot may still help — but its native advantage is weaker.

Enterprise AI assistants will be strongest when they sit inside the system where work already happens. Copilot is not just an assistant — it is an AI interface to the Microsoft work environment.

Claude Cowork: delegated work, not just chat

"Give Claude a goal, get back a finished deliverable."

Claude Cowork represents a different kind of AI product. It is not merely a chatbot, and it is not exactly a company wiki assistant. It is closer to delegated work.

The product lets users assign Claude a goal — and Claude works across files and applications to return a finished deliverable. That is the important phrase: finished deliverable. This changes the user's relationship with the model. In a normal chat, the user asks for an answer. In a delegated work system, the user assigns an outcome. The assistant may need to inspect files, use applications, gather information, make edits, and return something complete.

Claude Cowork is useful for tasks such as:

• Turning scattered notes into a structured memo.
• Preparing a client brief from files and context.
• Updating a document based on comments.
• Researching a market and producing a report.
• Cleaning up a messy folder of materials.
• Preparing meeting follow-ups.
• Drafting operational documents from existing files.
• Completing repetitive knowledge-work tasks that require context switching.

Claude Cowork is aimed at the messy middle of work. Not every task is a clean API call. Many tasks involve files, judgment, partial information, style, context, and synthesis. That is where delegated work agents become valuable.

This also fits the larger shift from execution to oversight: the user assigns the task, reviews the output, corrects direction, and decides what is acceptable — without needing to manually perform every step.

Perplexity: the research and answer layer

"Use when the bottleneck is finding, verifying, and synthesizing information."

Perplexity is the research and answer layer of the AI workbench — an AI-powered answer engine that provides accurate, trusted, real-time answers with citations. Perplexity Enterprise extends that to a secure platform for deep research and complex projects across files and tools.

This makes Perplexity different from a personal memory tool or an execution agent. It is strongest when the user needs to know what is true, what has changed, what sources say, and how to synthesize public or internal information quickly.

Perplexity is useful for:

• Market research and competitive analysis.
• Company intelligence and news monitoring.
• Technical research with cited sources.
• Industry scans and source-backed summaries.
• Quick explanations that need real-time freshness.
• Research workflows that need verifiable output.

Perplexity also fits into the workbench as a source of external context. Obsidian and Notion tell the agent what is true inside your world. Perplexity helps discover what is true outside your world. A personal or team memory system can only store what you already know — a research engine updates the system with new information.

The limitation is that research is not execution. Perplexity can help answer and synthesize, but it is not an agent that updates your CRM, sends emails, changes files, or moves tasks through a board. It is a powerful research surface, not the whole operating system.