L&#39;Atelier .

How we built teams of AI agents that don't loop

Handoff maps, three ways of working and hard budgets: the coordination runtime that makes a team of AI agents usable in a product, not just in a demo.

Our AI agents run on workflows, and that's deliberate

Pure agents decide every step at run time, which makes them flexible and hard to trust unattended. AWFlow's agents run on ordinary workflows instead: a deterministic skeleton with AI in the slots. Why we chose it and how it shows up in the code.

4 min de lecture15 octobre 2026

Recording clicks into replayable browser workflows: selectors, event mapping and MV3 gotchas

How a browser recorder turns clicks and typing into steps you can replay: capture-phase listeners, mapping raw events to intent, selectors that survive the next deploy, and a port that reconnects when the service worker sleeps.

4 min de lecture15 octobre 2026

Sharing an AI agent: what to keep, what to strip, and what never to trust

An agent is more than a prompt: it has memory, skills, rules and connections. Here is where we drew the line between the reusable setup and the personal state, and the rules we follow when someone installs a template.

11 min de lecture15 octobre 2026

We audited our own AI agent harness: 9 failure modes and the fixes

We audited the loop, context, tools, cancellation and failover around our in-browser AI assistant and found 21 gaps. Almost none were about the model. Here are the failure modes, the fixes, and a checklist for your own agent.

7 min de lecture15 octobre 2026

Workflows that survive a closed tab: durable waits with checkpoints and chrome.alarms

How a browser-extension workflow can pause for up to 30 days, survive a closed tab, a browser restart or an extension update, and resume where it stopped: checkpoints in storage, wake-ups owned by chrome.alarms.

Your browser is quietly an operating system: the 4 WebAssembly engines in our extension

One Chrome extension, no backend for its core features: a real database, two AI inference engines and a full Python interpreter, all running as WebAssembly. What each engine does, how it's wired, and what it costs.

Your context window is a budget, not a bucket

Most context-window bugs come from one assumption: the prompt is the chat history, so cut the oldest messages. In an agent, history is often the smallest part. Seven rules for budgeting a prompt, from a 128k cloud model down to a 2,048-token model in the browser.

3 min de lecture13 octobre 2026

LangChain and LangGraph in the browser: how an AI agent builds your workflow

An agent is not one big prompt. It is a graph of steps with branches, loops and a memory of where it paused. Here is how that works, with the real graph behind AWFlow's assistant.

2 min de lecture13 octobre 2026

Le problème des 5 onglets : laissez votre navigateur faire le copier-coller

Une petite tâche, cinq onglets : le prix d'un concurrent, une deuxième boutique, un tableur, un assistant de chat et votre boîte mail. Voici pourquoi cela finit par coûter cher, et comment un seul workflow dans votre navigateur s'en charge pour vous.

Why Your AI Agent Needs a Browser, Not Just an API

Most AI agents are built API-first, which leaves out the pages, forms and tools where real work happens. Why agents need a browser, how browser-native agents read and act on pages with your approval, and why it is better for privacy.

Build an AI Form-Filler That Works on Any Website

Job applications, support tickets, vendor portals: the same fields, the same answers. Here is how a workflow in your browser reads a form, picks the right value for each field and fills it, with you approving before anything is sent.

4 min de lecture27 avril 2026

How to Monitor Competitor Prices Automatically with AI

Stop checking competitor prices by hand every Monday. A browser workflow watches the price on any page, keeps its history, alerts you the day it drops and can write you an AI briefing: no code, no API needed.

Automate Your LinkedIn Prospecting Without Getting Banned

Most LinkedIn tools get accounts flagged because they act from outside your browser, at bot speed. What to automate, what never to automate, and a lower-risk research workflow that runs in your own session, at your pace.

5 Browser Tasks You’re Doing Manually That AI Can Do for You

Summarising pages, copying prices, filling forms, refreshing pages for changes, translating text: five browser rituals that eat your week, and the no-code workflow that takes each one off your hands, right in your browser.

Browser Automation for Non-Developers: A Practical Guide

Browser automation used to mean Python, Selenium and servers. Today you point at the page and connect blocks. What you can automate without code, how to build your first workflow step by step, and the myths to drop.

The 5-Tab Problem: How Knowledge Workers Lose 2 Hours a Day to Browser Switching

Data source, spreadsheet, AI assistant, chat, the actual task: five tabs you cycle through all day. What the switching costs, what the research says, and how one browser workflow collapses five tabs into one.

4 min de lecture27 avril 2026

How to Build a Personal AI Research Assistant That Runs in Your Browser

Stop copying pages into a chatbot tab. Build an AI research assistant that runs on the page you are reading: one hotkey to summarise, extract claims or answer questions, with a local model when the material is sensitive.

4 min de lecture27 avril 2026

What Is an Agentic Workflow? (And Why It’s Different from Regular Automation)

Rule-based automation follows the same steps every time and breaks on surprises. An agentic workflow reads the situation, reasons with AI and acts. What that means, how it differs, and why the browser is the place to run it.